awesome-ai-agents-2026

🚀 Start Here

New to AI agents? Follow this path: 1. 📖 Understand — what an agent actually is vs. a chatbot 2. 🗺️ Find your scenario → Scenario Guide 3. 🧩 Copy a proven setup → Stack Recipes 4. 🔍 Pick the right tool → Compare Tables 5. ⚠️ Avoid common mistakes → Anti-Picks

Already building? Jump to: - 🆕 Latest additions (August 2026) • 🛡️ Security • 💰 Cost comparison

Quick Navigation

Contents

• 🧠 Foundation Models 2026
• 🎨 Multimodal & Generative AI
• 🔗 Agent Protocols & Standards
• 🏗️ Agent Frameworks
• 🛠️ Agent IDEs & Visual Builders
• 🧠 Agent Memory
• 🔌 Tool & API Integration
• 💱 Agent Economy & Marketplaces
• 🧪 Agent Sandboxing & Compute Isolation
• 🛡️ Agent Security
• 🔍 RAG & Knowledge
• 💻 Coding Agents
• 🤖 Physical AI & Embodied Agents
• 🎮 Agent Simulation & World Models
• 📊 Benchmarks & Leaderboards
• 🖥️ Computer Use & Desktop Agents
• 🌐 Browser & Web Agents
• 🗣️ Voice & Multimodal Agents
• 📱 Personal AI Agents
• 📱 Mobile Agents
• 🏢 Enterprise Agent Platforms
• 📊 Agent Evaluation & Observability
• 🔬 AI Research Tools
• 📚 Learning Resources
• 🇨🇳 Chinese AI Ecosystem
• 📝 Compare — Side-by-Side Tables
• 🗺️ Scenario Guide — What Should I Use For…
• 📋 Stack Recipes — Curated Tool Combinations
• ⚠️ Anti-Picks — What NOT to Use For…
• 🌟 Notable Agent Projects of 2026
• 📅 2026 AI Timeline

🧠 Foundation Models 2026

The latest large language models powering the AI ecosystem, organized by company. 60+ models from 20+ providers.

OpenAI

• GPT-Live-1 / GPT-Live-1 mini - 🆕 July 8, 2026. OpenAI's full-duplex conversational voice model replacing Advanced Voice Mode. ChatGPT-only — not exposed as an API model; for programmatic realtime voice use gpt-realtime-2.1, and for streaming transcription gpt-live-transcribe ($0.017/min). Listens and speaks simultaneously (zero turn-taking lag), handles interruptions, delegates complex queries to GPT-5.5 in the background while keeping the conversation flowing. GPT-Live-1 is default for paid users (Go/Plus/Pro); GPT-Live-1 mini is default for free users. Includes real-time live translation. Available on iOS, Android, and web.
• OpenAI Astra - 🆕 ⚠️ Announced August 1, 2026 (public release date TBD). OpenAI’s next-generation model family announced in pre-release. An internal version reportedly resolved 10 unsolved mathematical and theoretical computer science problems in a single session (group theory, quantum complexity). 249-page manuscript + Lean 4 machine-verifiable certificates published publicly. ⚠️ No API access or public weights. August 7, 2026 update: OpenAI slowed Astra development after preliminary evaluations could not rule out a first-ever "Critical" cyber capability level; the White House confirmed OpenAI voluntarily informed the administration of the delay. Release timeline further pushed out.
• GPT-5.6 Sol - 🆕 July 9, 2026 (GA; limited preview from June 26). OpenAI's frontier flagship in the GPT-5.6 family — "Sol" is the most capable tier with advanced reasoning, coding, biology, and cybersecurity capabilities plus "max" reasoning and "ultra" sub-agent mode. Available on ChatGPT, Codex, and the OpenAI API. Launch was delayed briefly at US government request for a national-security review; rolled out to all users in stages following the trusted-partner preview. August 6, 2026 update: improved accuracy and consistency in ChatGPT; GPT-5.6 Luna expanded to free users for unlimited everyday chats. ⚡ August 13, 2026: new Ultrafast service tier (limited API preview) serves GPT-5.6 Sol at up to 14× Standard speed / ~750 output tokens per second, powered by Cerebras hardware.
• GPT-5.6 Terra - 🆕 July 9, 2026. Mid-tier model in the GPT-5.6 family offering GPT-5.5-parity performance at approximately 2× lower cost. Designed for cost-efficient production workloads.
• GPT-5.6 Luna - 🆕 July 9, 2026. The fastest and most cost-efficient tier of GPT-5.6 — optimised for high-volume, speed-critical tasks.
• ChatGPT Work - 🆕 July 9, 2026. OpenAI's agent that turns a goal into finished work — acts across connected apps and files, stays on a project for hours, creates slides/sheets/docs/web apps, runs scheduled tasks, and uses desktop computer-use with a built-in browser. Powered by GPT-5.6. Rolling out on web/mobile starting with Pro, Enterprise, and Edu (Plus/Business next); the desktop app is available globally on Mac and Windows for all plans, including Free.
• Sites for ChatGPT - 🆕 June 2026. A Codex-powered ChatGPT feature that transforms plans and analyses into interactive, sharable websites and lightweight apps. In public beta as of the July 9, 2026 GPT-5.6 / ChatGPT Work launch.
• Codex Business Plugins - 🆕 June 2026. Enterprise enhancements bringing sales, data analytics, and creative production plugins directly to Codex.

• GPT-Rosalind - 🆕 June 3, 2026. Major update to OpenAI's life-sciences frontier model — stronger drug discovery, genomics, quantitative biology, and wet-lab troubleshooting (≈31% fewer tokens than GPT-5.5 on long-horizon genomics analyses). Research preview opened to eligible organizations worldwide; Novo Nordisk joins earlier partners Amgen, Moderna, the Allen Institute, and Thermo Fisher.

• GPT-5.5 - 🆕 Released April 23, 2026 (codename "Spud"). OpenAI's new frontier model for agentic tasks: coding, online research, data analysis, autonomous tool navigation. Significant gains in reasoning, consistency, and long-horizon task handling. Available on ChatGPT Plus / Pro / Business / Enterprise.

• GPT-5.5 Pro - 🆕 April 23, 2026. Parallel test-time compute variant for higher-accuracy cognitive tasks. Pro / Business / Enterprise tiers.
• GPT-5.5 Instant - 🆕 May 5, 2026. New ChatGPT default model. Efficiency-first upgrade with ~50% lower hallucination rate on high-stakes prompts; available on free tier.
• GPT-5.5-Cyber - 🆕 April 30, 2026. Cybersecurity-specialized variant of GPT-5.5, rolled out via OpenAI's Trusted Access for Cyber (TAC) program to vetted defenders, government, critical infrastructure operators, and security vendors. Not available to the general public.
• OpenAI Daybreak - 🆕 May 12, 2026. Cyber-defense platform bundling GPT-5.5 + GPT-5.5-Cyber + Trusted-Access-for-Cyber for AI-powered vulnerability detection and patch validation; preview access extended to EU governments and security vendors.
• GPT-5.6-Cyber - 🆕 ⚡ August 10, 2026. OpenAI's cybersecurity-specialized model built on GPT-5.6 Sol, made available through the Daybreak Red vetted program for authorized vulnerability research and exploit validation. Capable of identifying zero-day vulnerabilities and developing exploit chains; access limited to cleared security professionals. Daybreak also expanded to AWS Bedrock on August 11.
• GPT-Realtime-2 - 🆕 May 8, 2026. GPT-5-class reasoning brought to the Realtime API, 128K context, parallel tool calls with audio feedback, adjustable reasoning effort.
• GPT-Realtime-Translate - 🆕 May 8, 2026. Live speech-to-speech translation across 70+ input languages and 13 output languages.
• GPT-Realtime-Whisper - 🆕 May 8, 2026. Streaming low-latency speech-to-text companion to GPT-Realtime-2.
• OpenAI Deployment Company (DeployCo) - 🆕 May 11, 2026. New OpenAI-majority-owned services entity for enterprise AI rollout. Backed by $4B+ from TPG / Advent / Bain Capital / Brookfield / Goldman Sachs / SoftBank and consulting partners Bain & Company, Capgemini, McKinsey. Built around Forward Deployed Engineers; absorbs the Tomoro AI consulting acquisition (~150 engineers).
• Codex on Mobile - 🆕 May 14, 2026. ChatGPT iOS/Android can now remote-control the Codex desktop app — review outputs, approve actions, switch models, and kick off new tasks from the phone while the live session runs on Mac (Windows next). Rolling out as preview to Free, Plus and Go users.
• OpenAI ↔ Malta partnership - 🆕 May 16, 2026. First country-wide deal: every Maltese citizen / resident aged 14+ gets a free 1-year ChatGPT Plus subscription after completing a 2-hour AI literacy course built by the University of Malta. Part of the "OpenAI for Countries" initiative; phased rollout starting May 2026.
• OpenAI ↔ Dell Codex partnership - 🆕 May 18, 2026. Brings Codex to hybrid and on-premises enterprise environments via Dell Technologies infrastructure — first major Codex distribution channel outside the public cloud, targeted at regulated industries needing data-residency control.
• ChatGPT Safety Updates — sensitive-conversation tracking - 🆕 May 18, 2026. ChatGPT's safety systems updated to detect and track subtle escalation cues across long sessions for acute risks (suicide / self-harm / harm to others), with cross-session state retention.
• OpenAI Guaranteed Capacity (Compute Annual Pass) - 🆕 May 19, 2026. Long-term compute reservation product for enterprise AI products / agents / workflows. 1, 2, or 3-year terms; longer terms unlock larger discounts. OpenAI's structural response to the Anthropic "Priority Tier" model.
• OpenAI ↔ Google SynthID + C2PA content provenance - 🆕 May 19, 2026. OpenAI partners with Google to add durable cross-platform SynthID watermarking to ChatGPT/Sora images, joins C2PA, and previews a public "is-this-image-from-OpenAI" verifier. First major frontier-lab interop on watermarking.
• GPT-5.4 - Released March 2026. Frontier model with 1M-token context, advanced coding, computer use, tool search. BenchLM 94, SWE-bench Verified 77.2%, OSWorld 75% (beats human).
• GPT-5.4 Pro - Higher-accuracy variant of GPT-5.4. BenchLM 92.
• GPT-5.3 - Early 2026. Includes GPT-5.3 Instant (conversations) and GPT-5.3-Codex (coding).
• GPT-5.2 - Released Dec 2025. State-of-the-art reasoning, long-context understanding, and vision.
• GPT-5 - Launched August 2025. The default model in ChatGPT, replacing GPT-4o. Multimodal with variants: gpt-5, gpt-5-mini, gpt-5-nano.
• GPT-4o - Omni model with native text, vision, and audio. Retired from ChatGPT Feb 2026 but still available via API.
• GPT-4.5 - 📦 Retired from ChatGPT late June 2026 (API access continues; conversations auto-migrated to GPT-5.5). Released Feb 2025 as a research preview — the last GPT-4-family model in ChatGPT. o3 retiring from ChatGPT Aug 26, 2026.
• o3 / o4-mini - Reasoning models with chain-of-thought for complex problem solving. Released April 2025. o3 leaves ChatGPT on Aug 26, 2026; the o3-2025-04-16 and o3-pro-2025-06-10 API snapshots are removed on Dec 11, 2026, with gpt-5.6-sol named as the replacement (deprecations).
• Codex CLI - Open-source terminal-based coding agent powered by OpenAI models.

Anthropic

• Claude text watermarking + content credentials - 🆕 ⚡ August 14, 2026. Anthropic adds invisible SynthID-Text-based watermarking (Google DeepMind's method) to future Claude models globally at launch, plus C2PA content credentials on generated images/files (.png/.jpg/.svg); models released before August 2, 2026 get it "over the coming months", and a detection API is coming. Implemented to comply with the EU AI Act after Anthropic signed the EU transparency Code of Practice in July. Anthropic says watermarked text is indistinguishable to readers.
• Claude Opus 5 - 🆕 July 24, 2026. Anthropic's fifth-generation flagship — nears Fable 5 performance at a significantly lower price ($5/$25 per million input/output tokens). 1M-token context window, 128K output tokens. Now the default model on Claude Max. API: claude-opus-5. Available on Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI.
• Claude Fable 5 (Global Reinstatement) - 🆕 July 1, 2026. After US Commerce Department export controls were lifted on June 30, Anthropic reinstated global access to Fable 5 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. A new safety classifier blocking the Amazon-discovered jailbreak was deployed (blocks the reported behavior in >99% of cases). Pro/Max/Team and select Enterprise plans got Fable 5 included for up to 50% of weekly usage through July 7, then via usage credits; cloud re-enablement on AWS, Google Cloud, and Microsoft Foundry to follow. Mythos 5 remains restricted to vetted US entities.
• Claude Sonnet 5 - 🆕 June 30, 2026. The most agentic Sonnet yet — planning, browser/terminal tool use, and autonomous operation at a level that recently required Opus-class models. Performance approaches Opus 4.8 on agentic search (BrowseComp) and computer use (OSWorld-Verified) at higher effort settings, with a much wider cost-performance range than Sonnet 4.6. Now the default model for Claude.ai Free/Pro; also on Max/Team/Enterprise, Claude Code, and the API as claude-sonnet-5. August 10, 2026 update: the introductory $2/$10 per million input/output pricing was made permanent — the previously scheduled increase to $3/$15 on September 1 will not occur. Anthropic reports a lower rate of undesirable behaviors than Sonnet 4.6.
• Claude Fable 5 - 🆕 June 9, 2026. Anthropic's first publicly available Mythos-class model — a capability tier above Opus. Surpasses Opus 4.8 across software engineering, knowledge work, vision, and scientific research benchmarks. Ships with built-in safeguards (sensitive cyber/bio queries may be rerouted to Opus 4.8). $10 / $50 per million in/out tokens. Available via Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. ⚠️ Access suspended June 12, 2026 — a US government export-control directive ordered Anthropic to disable Fable 5 and Mythos 5 for all customers pending security review. ✅ Export controls lifted June 30, 2026; access restored July 1 with a new cybersecurity classifier — see entry above (statement).
• Claude Mythos 5 - 🆕 June 9, 2026. The same underlying Mythos-class model as Fable 5 with fewer restrictions, deployed only to vetted partners (cybersecurity firms, infrastructure providers) through Project Glasswing in collaboration with the US government. Successor to the April Claude Mythos Preview. ⚠️ Suspended June 12, 2026 alongside Fable 5 under a US export-control directive. ✅ Partially reinstated June 26, 2026 — US Commerce Secretary Lutnick restored access to 100+ approved US companies and federal agencies; broader reinstatement ongoing (statement).
• Claude Opus 4.8 - 🆕 May 28, 2026. Major Opus refresh: codebase-scale migrations, sharper agentic judgment, dynamic workflows research preview with hundreds of parallel sub-agents in a single session, manual effort-control panel, 3× cheaper Fast mode at the same $5 / $25 per million in/out. Available on Anthropic native + Amazon Bedrock + AWS Claude Platform + Google Cloud + Microsoft Foundry. Teases an upcoming Mythos-class model series for limited orgs.
• Claude Opus 4.7 - 🆕 Released April 16, 2026. Advanced software engineering (SWE-bench Verified 87.6%), enhanced vision, proactive code verification. Supports /think xhigh reasoning effort. 1M-token context.
• Claude Opus 4.6 - Released Feb 2026. 1M-token context, 14.5-hour task horizon. Leads Arena chat leaderboard.
• Claude Sonnet 4.6 - Released Feb 2026. Frontier coding and agentic performance, 1M token context window.
• Claude Mythos Preview - 🆕 April 2026 gated research preview. BenchLM 99 (top of leaderboard), SWE-bench Verified 93.9%. Limited to Project Glasswing partners.
• Claude Opus 4 - Released May 2025. Advanced reasoning and complex task execution.
• Claude Sonnet 4 - Released May 2025. Balanced performance and cost for a wide range of tasks.
• Claude Code - Agentic coding tool operating directly in your terminal. Powered by Opus 4.7 with /think xhigh support. July 2026: desktop app gains a built-in browser enabling live website interaction (scraping, debugging, live-page inspection); Fable 5 model available since July 1.
• Claude Security - 🆕 May 1, 2026. Public beta. Enterprise security tool powered by Opus 4.7 — scans entire codebases for vulnerabilities and generates targeted patches with confidence rating, severity, reproduction steps, and recommended fixes. Available to Enterprise customers via claude.ai/security.
• Claude Finance Agents - 🆕 May 5, 2026. Ten Opus-4.7-powered specialised agents for pitchbook authoring, KYC, month-end close, deal screening, etc. Deployable as Claude Cowork plugins, Claude Code skills, or Managed-Agents cookbooks.
• Claude Finance JV - 🆕 May 4, 2026. $1.5B Claude deployment joint venture with Goldman Sachs and Blackstone embedding Anthropic engineers in mid-market Wall Street firms.
• Claude Add-ins / Dreaming / Outcomes / Multi-agent orchestration - 🆕 May 8, 2026 (Code with Claude 2026). Anthropic introduces Add-ins, scheduled memory review between sessions ("Dreaming"), rubric-driven "Outcomes", and a lead-agent + sub-agent orchestration model with shared filesystem and auditable trace.
• Anthropic ↔ SpaceX Colossus 1 - 🆕 May 6, 2026. Anthropic takes all available capacity at SpaceX's Colossus 1 Memphis datacenter (>220K NVIDIA H100/H200/GB200 GPUs, 300+ MW) for Claude Opus inference. Doubles Claude Code 5-hour rate limits on Pro/Max/Team/Enterprise; also lifts peak-hour limits.
• Anthropic ↔ AMD (up to 2 GW of Instinct MI450) - 🆕 July 22, 2026. Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series (MI455X) GPUs in AMD Helios rack-scale systems with EPYC "Venice" CPUs, Pensando networking and ROCm; the first gigawatt begins in H1 2027. AMD has committed a strategic equity investment of up to $5 billion in Anthropic, plus a multi-year engineering collaboration. Builds on Anthropic's existing MI355X usage — a deliberate hardware-diversification move alongside its TPU, Trainium and SpaceX Colossus capacity.
• Anthropic's position on open-weights models - 🆕 July 27, 2026. Dario Amodei responds to reports that US officials are weighing a ban on Chinese open-weights models: "Anthropic has never advocated for a ban on open-weights models." He calls non-dangerous open weights "a public good" and instead backs chip export controls plus a smuggling crackdown, deterrence of industrial-scale distillation, and mandatory pre-release safety testing for all sufficiently capable models, open and closed. Useful primary source for anyone tracking the 2026 open-vs-closed policy fight.
• Claude for Legal - 🆕 May 12, 2026. New legal stack on top of Claude Cowork: 20+ MCP connectors (iManage, NetDocuments, DocuSign, Ironclad, LexisNexis, Westlaw, Harvey, Everlaw, Relativity, CourtListener…) + 12 practice-area plugins (commercial, employment, privacy, product, corporate, AI governance, litigation associate, law-student bar-exam). Microsoft Word / Outlook / Excel / PowerPoint orchestration built in.
• Claude for Small Business - 🆕 May 13, 2026. Small-business toggle inside Claude Cowork — 15 pre-built agentic workflows across finance / ops / sales / marketing / HR / customer service, native connectors for QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365. Bundled with a free PayPal-backed "AI Fluency for Small Business" course and a 10-city US workshop tour kicking off in Chicago.
• Anthropic ↔ Gates Foundation $200M - 🆕 May 14, 2026. 4-year, $200M partnership pairing grants + Claude usage credits + Anthropic engineers on global-health, life-sciences, education, and agriculture programs. All tools produced under the program will be freely available; first focus areas include vaccine R&D for polio / HPV / preeclampsia and agriculture-specific Claude extensions.
• Anthropic ↔ PwC strategic alliance expansion - 🆕 May 14, 2026. PwC commits to global rollout of Claude Code + Claude Cowork, certifies 30,000 PwC professionals, and stands up a joint "Agentic Enterprise" Center of Excellence — focused on agentic build, AI-native deals, and finance / supply-chain / HR reinvention.
• Anthropic ↔ Financial Stability Board briefing (Claude Mythos) - 🆕 May 18, 2026. Anthropic briefs the global FSB on Claude Mythos cyber-flaw discovery capabilities — first time a frontier lab briefs a G20-level financial-stability regulator on a frontier model's offensive-security implications.
• Code with Claude 2026 sessions on YouTube - 🆕 May 18, 2026 (sessions published). Full developer-conference recordings (May 6 event) go public: Claude Code roadmap, Claude Developer Platform updates, Managed Agents dreaming + multi-agent orchestration, and partner deployments.
• Widening the conversation on frontier AI - 🆕 May 19, 2026. Anthropic publishes its framework for engaging diverse traditions (religious, philosophical, indigenous) in frontier-AI safety dialogue. Companion to ongoing public-engagement work.
• Bristol Myers Squibb ↔ Anthropic Claude Enterprise - 🆕 May 20, 2026. BMS adopts Claude Enterprise as its shared intelligence platform for 30,000+ employees globally, embedding agentic Claude into drug-discovery / development / delivery workflows. First top-5 pharma enterprise-wide Claude deployment.

Google DeepMind

• Gemini 3.7 Flash - 🆕 ⚡ August 13, 2026. Google's new "most intelligent workhorse model" for coding and agents, shipped just three weeks after 3.6 Flash — and before the still-missing 3.5 Pro. FrontierCode 1.1 43.6% (vs 34.4% for 3.6 Flash), DeepSWE v1.1 65.3% (vs 49.0%). Introductory pricing $0.75/$3.75 per million in/out through Dec 31, 2026 (then $1.50/$7.50). Available in AI Studio, Android Studio, Antigravity, and the Gemini Enterprise Agent Platform; powers Gemini Spark for AI Pro/Ultra subscribers.
• Gemini 3.6 Flash - 🆕 July 21, 2026. Google's Flash tier — stronger on complex agentic and multimodal tasks while using fewer tokens, at a lower price point than 3.5 Flash. API id gemini-3.6-flash. Documented in the official Gemini API cookbook alongside thinking-mode guides. Superseded as the top Flash tier by 3.7 Flash on August 13, 2026.
• Gemini 3.5 Flash-Lite - 🆕 July 21, 2026. The fastest, lowest-cost model in the 3.5 family; outperforms prior Flash-Lite generations for high-throughput execution. API id gemini-3.5-flash-lite. Now the cheapest Gemini tier, superseding 3.1 Flash-Lite for new builds.
• Gemini 3.1 Pro (preview) - Google's most capable Gemini as of late July 2026, served as gemini-3.1-pro-preview. GPQA Diamond 94.3% (world-record at launch), ARC-AGI-2 77.1%, BenchLM 94. ⚠️ Still carries the -preview suffix and has no free tier.
• Gemini 3.5 Pro - ⚠️ Delayed — still unreleased as of August 13, 2026 (limited enterprise preview with partners ongoing). Google's upcoming flagship with a reported 2-million-token context window and Deep Think reasoning mode; substantially improved coding and agentic workflow capabilities. Announced at Google I/O in May 2026 for a June release, then postponed after Google scrapped and rebuilt the base model over disappointing coding performance. At the Gemini 3.7 Flash launch (Aug 13) Google declined to discuss 3.5 Pro's fate and confirmed it is now training Gemini 4. Competes directly with GPT-5.6 Sol and Claude Fable 5.
• Gemma 4 12B - 🆕 June 2026. Novel multimodal open model with a unified, encoder-free architecture processing text, images, and audio in a single pass. Runs locally on 16 GB VRAM.

• DiffusionGemma - 🆕 June 2026. 26B MoE open model using text-diffusion for up to 4× faster generation than autoregressive models.

• Gemini 3.5 Flash - 🆕 May 19, 2026 — Google I/O 2026. Default model powering the Gemini app and Google Search AI Mode. Marketed as ~4× faster than other frontier models in output tokens/sec while outperforming Gemini 3.1 Pro on key benchmarks. Gemini 3.5 Pro was slated for June 2026 but has been delayed (see above).

• Gemini Omni / Omni Flash - 🆕 May 19, 2026 — Google I/O 2026. New Google DeepMind multimodal world-model family aimed at AGI. Omni Flash, the first shipped variant, can take any input modality and generate any output (starting with video; image and text generation following). Direct lineage to Gemini Robotics / Genie line of work.
• Gemini 3.1 Pro - Released Feb 2026. BenchLM 94, GPQA Diamond 94.3% (world-record), ARC AGI2 77.1%. $2/1M tokens flagship.
• Gemini 3.1 Flash Live - 🆕 April 2026. Real-time multimodal streaming for voice assistants and interactive agents. Low latency, long context.
• Gemini 3.1 Flash-Lite (GA) - 🆕 May 8, 2026. Generally available on Gemini API / AI Studio / Vertex AI. Fastest and most cost-efficient model in the Gemini 3 family — built for low-latency code completion, real-time UX, and agentic developer tools; matches Gemini 2.5 Flash quality at significantly lower cost.
• Gemini Omni Flash — voice-controlled video editing rollout - 🆕 May 28, 2026. Omni Flash starts rolling out to consumers via the Gemini app, Google Flow, and YouTube Shorts as the editing engine — conversational cinematic zooms / background swaps / weather edits driven by text, voice, image, or audio prompts; no traditional NLE required.
• Gemini Spark (24/7 personal AI agent) - 🆕 May 19, 2026 — Google I/O 2026. Cloud-resident personal AI agent that runs 24/7 on user intent, integrates Gmail / Chat first, then ~30+ third-party tools via MCP (Adobe / Dropbox / Uber). Available to Google AI Ultra subscribers in the US within the I/O week.
• Google AI Ultra ($100/month tier) - 🆕 May 19, 2026 — Google I/O 2026. New top consumer subscription targeted at developers / creators / power users. Gates Gemini Spark, highest Gemini 3.5 quotas, and the upcoming Gemini 3.5 Pro.
• Gemini 3.1 Flash / Flash Lite - Fast, cost-efficient models for high-throughput applications.
• Gemini 4 (Open) - 🆕 Released April 2026. Open model family: 2B / 4B / 26B / 31B variants. Strong science reasoning and document understanding, local deployment ready.
• Gemini 2.5 Pro / Flash - GA June 2025. Thinking model with 1M context.
• Gemma 4 31B - 🆕 April 2026. GPQA Diamond 84.3%. Strong open-weight alternative for on-device reasoning.
• Gemma 3 - Previous open model family for on-device and research use.
• Gemini Robotics ER-1.6 - 🆕 April 14, 2026. Upgraded robotics AI with improved spatial and physical reasoning. Partnership with Agile Robotics for real-world deployment.

Meta

• Muse Image - 🆕 July 7, 2026. Meta Superintelligence Labs' most advanced image generation model to date — an "agentic" image model that performs intermediate reasoning steps (web search, code execution, self-refinement) before producing high-quality visuals. Integrated into the Meta AI app, Instagram Stories (US), and WhatsApp in limited countries (Facebook coming soon). Note: a controversial feature allowing images from other users' public Instagram profiles was added then removed on July 10 after feedback.
• Muse Spark 1.1 - 🆕 July 9, 2026. Multimodal reasoning model designed for agentic tasks from Meta Superintelligence Labs — available through a new public preview of the Meta Model API. Marks a strategic shift toward proprietary revenue-focused models alongside Meta's open-source Llama line.
• Muse Video - 🆕 July 7, 2026 (preview). Video generation model from Meta Superintelligence Labs, built on the same foundational technology as Muse Image; ranks #3 on Arena for text-to-video. Previewed alongside the Muse Image launch — "coming soon to creators and Meta AI."
• Muse Spark 1.2 + Muse Code (beta) - 🆕 ⚡ August 5, 2026. Muse Code is MSL's terminal coding agent (async background agents, replay-exact local event log, /plan / /grill / /goal skills), powered by the new Muse Spark 1.2 — trained on whole-repository generation and long-horizon coding. Spark 1.2 is also available in the Meta Model API with expanded global access.
• Muse Glimmer 30B - 🆕 ⚡ August 10, 2026. Open-weight, 30B-parameter multimodal model from Meta Superintelligence Labs, Apache 2.0 license. Designed for always-on local agent workflows and optimized to run on a single consumer GPU or Apple Silicon. 131K-token context window, 100+ language training, DFlash acceleration. Distilled from Muse Spark; tuned for coding, evaluation, and agentic tasks. Available on Hugging Face with llama.cpp / MLX / ExecuTorch integrations.
• Llama 5 — ❌ Does not exist. Removed from this list on 2026-07-30 after verification. A "Llama 5, 600B+, April 8 2026" entry circulated widely in AI-news aggregators and LLM search summaries, and was previously listed here. It does not hold up: the meta-llama Hugging Face organisation contains no Llama-5 weights of any kind (newest Llama-family upload is Llama-4-Maverick, May 2025), and Wikipedia's Llama article states "the latest version is Llama 4, released in April 2025" and that Muse Spark replaced the Llama line in April 2026. Treat any "Llama 5" claim as unverified until Meta publishes weights or a newsroom post. See Muse Spark above for what actually shipped.
• Muse Spark - 🆕 April 9, 2026. First model from Meta Superintelligence Labs (MSL). Natively multimodal reasoning model powering Meta AI app, smart glasses, and features across Facebook / Instagram / WhatsApp / Messenger.
• Llama 4 Scout - 109B total params (17B active), MoE with 16 experts, 10M token context window, multimodal. Runs on single H100.
• Llama 4 Maverick - 400B total params (17B active), 128 experts, 1M context. Outperforms GPT-4o on multimodal benchmarks.
• Llama 4 Behemoth - 2T parameters (288B active). In training — Meta's frontier model rivaling top closed-source models.
• Llama 3.3 70B - Strong instruction following and reasoning, open-weight under Llama Community License.

Sakana AI

• Sakana RL Conductor - 🆕 Paper April 27, 2026; Fugu beta late-April / early-May 2026. 7B RL-trained orchestrator (built on Qwen2.5-7B) that routes subtasks between GPT-5, Claude Sonnet 4, Gemini 2.5 Pro, etc. SOTA on LiveCodeBench (83.9%) and GPQA-Diamond (87.5%) at ~1.8K tokens/query — roughly 6× cheaper than other multi-agent ensembles.
• Sakana Fugu - 🆕 Beta April 24-25, 2026. Commercial multi-agent orchestration service productising the RL Conductor research. OpenAI-compatible API with two tiers: Fugu Mini (low-latency) and Fugu Ultra (max performance); strong reported results on SWE-Pro, GPQA-D and ALE-Bench.

Zyphra

• ZAYA1-8B - 🆕 May 6, 2026. MoE reasoning model (<1B active) trained end-to-end on AMD Instinct MI300X clusters. Apache 2.0 weights on Hugging Face + serverless endpoint on Zyphra Cloud; aimed at math, code, and dense reasoning per active parameter.
• ZAYA1-8B-Diffusion-Preview - 🆕 May 14, 2026. First MoE diffusion language model converted from an autoregressive LLM and the first diffusion LM trained on AMD GPUs. Generates 16 tokens per step, achieving up to 7.7× inference speedup vs the autoregressive base. Built with Zyphra's TiDAR recipe + CCA attention.

Thinking Machines Lab

• Inkling - 🆕 July 15, 2026. Founded by Mira Murati (former OpenAI CTO). 975B MoE parameters (41B active), pretrained on 45T tokens, 1M-token context window. Natively multimodal — text, image, audio, and video in a single model. Apache 2.0 open weights on Hugging Face. Also ships Inkling-Small (12B active parameters). Available via the Thinking Machines API and Hugging Face Inference.
• Inkling-Small - 🆕 July 30, 2026 (weights released). Compact variant of Inkling — 276B total / 12B active, same native multimodal architecture (text/image/audio), 1M-token context, Apache 2.0. Scores 31.6% on HLE text benchmark — slightly outperforming the larger 975B Inkling (29.7%) on that metric, validating the efficiency-first design. Available via Thinking Machines API and Hugging Face.

Mistral AI

• Mistral Large 3 - 675B total / 41B active parameters, MoE, 256K context. Flagship open-weight multimodal model. Released Dec 2025.
• Mistral Medium 3.1 - Frontier-class dense model for enterprise. Multimodal, 128K context, 80+ coding languages. Released Aug 2025.
• Mistral Small 4 - 🆕 Released March 2026. 119B total / 6B active. Hybrid model combining reasoning, multimodal, and coding strengths.
• Magistral 1.2 - 🆕 2026 reasoning family challenging o3/o4-mini. Transparent and multilingual reasoning.
• Devstral 2 - 🆕 2026 agentic coding model. Best open-source model for coding agents.
• Codestral - 22B code generation model, 80+ programming languages, 32K context. Released May 2024.
• Pixtral Large - 124B multimodal model with 1B vision encoder, 128K context, processes 30+ high-res images.
• Ministral 3B/8B/14B - Compact models optimized for edge deployment and efficiency.
• Mistral Forge - 🆕 March 2026 platform for training custom LLMs on proprietary data.
• Mistral Medium 3.5 - 🆕 April 28, 2026. Dense 128B open-weight model, 256K context, Modified MIT license. Unifies instruction-following, reasoning, and coding.
• Leanstral 1.5 - 🆕 July 2, 2026. Formal-verification model for proof engineering in Lean 4 — 119B total / 6B active parameters, Apache 2.0, weights on Hugging Face plus a free API endpoint. Scores 100% on miniF2F, solves 587/672 PutnamBench problems, and discovered 5 previously unreported bugs across 57 real-world repositories.
• Robostral Navigate - 🆕 July 8, 2026. Mistral's first robotics model — an 8B embodied-navigation model that steers wheeled, legged, and flying robots through offices, homes, and outdoor spaces from natural-language instructions using only a single RGB camera (76.6% success rate on unseen validation). Trained fully in-house on ~400K simulated trajectories.
• Voxtral TTS - 🆕 March 26, 2026. 4B-parameter open-weight TTS built on Ministral 3B; multilingual, optimised for voice agents.

DeepSeek

• DeepSeek-V4-Pro-0813 (GA) - 🆕 ⚡ August 13, 2026. GA release of deepseek-v4-pro on app/web/API with "major Agent upgrades with strong production gains": selectable reasoning effort (low/high/max — also on V4-Flash), native OpenAI Responses API support optimized for Codex, and an "Expert Mode" in the app/web. ⚠️ Pricing changes August 16, 2026 (16:00 UTC): flat pricing ends and peak/off-peak tiering begins (peak 01:00–04:00 & 06:00–10:00 UTC; off-peak 50% cheaper). V4-Pro peak: $0.044 cache-hit / $1.32 cache-miss input, $3.96 output per 1M tokens — a significant increase (pricing).
• DeepSeek-V4-Pro - 🆕 April 24, 2026 (preview); production launch mid-July 2026. 1.6T total / 49B active MoE, 1M-token context. MIT license. Leadership in agent capabilities, world knowledge, reasoning; tops open-source benchmarks. 384K max output, 500-request concurrency. deepseek-v4-pro / deepseek-v4-flash are the production API models (V4-Pro serves the 0813 checkpoint since August 13 — see above; tiered peak/off-peak pricing from August 16, 2026).
• DeepSeek-V4-Flash - 🆕 April 24, 2026. 284B total / 13B active MoE, 1M context. MIT. Cost-efficient tier — from August 16, 2026: peak $0.014 cache-hit / $0.44 cache-miss input, $1.32 output, off-peak $0.007 / $0.22 / $0.66 per 1M tokens; 384K max output, 2,500-request concurrency (pricing).
• DeepSeek-V4-Flash-0731 - 🆕 July 31, 2026. Updated Flash checkpoint with enhanced agentic capabilities — same 284B/13B-active MoE architecture, same pricing/API model ID, but outperforms V4-Pro (Preview) on agent task benchmarks. Open weights on Hugging Face under MIT license. Drop-in replacement for deepseek-v4-flash API users.
• DeepSeek Agent Harness team - 🆕 May 19, 2026. DeepSeek hires a former Jane Street engineer to lead a new "AI harness" team building the deterministic scaffolding that turns DeepSeek V4 into autonomous, revenue-generating agents — first major signal DeepSeek is moving past raw-model R&D into agentic productisation.
• DeepSeek-V3.2 - Released Dec 2025. Advanced MoE architecture with 671B total parameters. V3.2 Speciale variant for enhanced reasoning. ⚠️ API model IDs deepseek-chat / deepseek-reasoner (V3.2-era) deprecated effective July 24, 2026 — superseded by V4-Flash modes.
• DeepSeek-R2 - 🧪 Unreleased/rumored. No official announcement, model card, or API ID exists as of mid-July 2026; reasoning is served via V4's Thinking mode.
• DeepSeek-R1 - Reasoning-focused model with chain-of-thought capabilities. Released Jan 2025.
• DeepSeek-Coder-V2 - Code generation model competitive with GPT-4 on coding benchmarks.

Alibaba (Qwen)

• Qwen3.8-27B - 🆕 ⚡ August 14, 2026. Open-weight 27B multimodal (text/image/video input) distillation of Qwen3.8-Max, released on Hugging Face under Apache 2.0 — sized for ~24 GB-VRAM consumer GPUs (RTX 4090-class). The promised open-weight companion to the Qwen3.8-Max launch, shipped on schedule.
• Qwen3.8-Max - 🆕 ⚡ August 3, 2026. Alibaba's new flagship — 2.4T-parameter MoE (95B active), 1M-token context, accepts text/image/video inputs. Designed for real-world work, research, and autonomous coding; Alibaba reports an internal run where the model spent 16 days building and refining a coding tool without human intervention. API via QwenCloud/Alibaba Cloud Model Studio ($2/M input, $6/M output). Open weights for the full model appeared on Hugging Face ~August 12 as Qwen/Qwen3.8-2.4T-A95B, with the 27B distillation following August 14. Companion QwenWork enterprise platform entered public beta Aug 2.
• Qwen3.7-Max - 🆕 May 20, 2026 — Alibaba Cloud Summit Hangzhou. New Qwen flagship purpose-built as the foundation for AI agents: agentic coding, complex reasoning, and long-horizon multi-step missions with sustained decision-making. Released alongside a full-stack AI infrastructure upgrade and new T-Head Zhenwu M890 AI accelerator chip. Worldwide developer/enterprise availability rolling.
• Qwen3.7-Max-Preview / Qwen3.7-Plus-Preview - 🆕 May 18, 2026. Preview ladder before the Hangzhou unveil. Ranked the highest of any Chinese model on LM Arena in both text and vision; sustained 1M-context evaluations.
• Qwen3.6-27B - 🆕 April 22, 2026. Dense 27B multimodal. Open-sourced. Focus: agentic coding + thinking-context preservation.
• Qwen3.6-Max-Preview - 🆕 April 18, 2026. Proprietary frontier preview. High coding/reasoning performance, 1M context window. Top-tier among Chinese models on coding benchmarks.
• Qwen3.6-35B-A3B - 🆕 April 15, 2026. MoE, 35B total / 3B active. Apache 2.0. Stability and real-world utility improvements.
• Qwen3.6-Plus - 🆕 April 2, 2026. Proprietary flagship. High value-per-token general model. Strong long-context, tool-calling, agentic behavior.
• HappyHorse 1.1 - 🆕 June 23, 2026. Alibaba's video-generation model (T2V/I2V/S2V, up to 15s 1080p with synced audio, strong multi-shot character consistency). HappyHorse 1.0 entered limited beta April 28, 2026 after launching anonymously and topping video leaderboards.
• Qwen3.5 Max Pro - April 2026. High-performance flagship. Enhanced coding and math reasoning, long context.
• Qwen3.5 Omni Plus - April 2026. Proprietary full-modal foundation model unifying text and image input.
• Qwen3-Max-Thinking - Alibaba's strongest thinking model. 1T+ parameters, enhanced agentic capabilities.
• Qwen3.5-Omni - March 2026. Fully omni-modal: language, vision, sound, motion. Speech recognition in 113 languages, 256K context.
• Qwen3-Coder-Next - Feb 2026. Open-weight coding agent model, MoE 80B total / 3B active.
• Qwen3 235B-A22B - MoE with dual-mode reasoning. Strong math, code, and commonsense reasoning.
• Qwen2.5 Coder 32B - Top open-source coding model.

xAI / SpaceXAI (Grok)

• Grok 4.6 - 🆕 ⚡ August 12, 2026. SpaceXAI's new frontier model for coding, agentic tasks, and knowledge work — built on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive/visual work, plus improved self-testing and verification. 500K-token context, text + image input, same $2/$6 per million in/out pricing (Fast variant at 2×). Matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index (61). Live in the xAI API, Cursor (new default), Grok Build, OpenRouter, Vercel, and Cloudflare; landed in GitHub Copilot August 14. (xAI completed its rebrand to SpaceXAI in July 2026 after SpaceX absorbed xAI in February.)
• Grok Bot - 🆕 ⚡ August 11, 2026 (early beta). Durable AI teammates that work on a persistent cloud computer, with messaging, approvals, connectors, and routines — xAI's entry into always-on autonomous agents. Available via SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium.
• Grok 4.5 - 🆕 July 8, 2026. Optimised for coding and agentic tasks through joint training with Cursor using real developer interaction data. Features a 500K-token context window, function calling, structured outputs, web/X search, code execution, document search, and context compaction. Priced at $2/$6 per million in/out tokens. EU API-console availability arrived July 17, 2026. Superseded as flagship by Grok 4.6 on August 12, 2026.
• Grok 4.3 GA - 🆕 May 2026. Grok 4.3 reached general availability on Microsoft Foundry and OCI Generative AI; xAI's flagship for agentic workloads with improved tool-calling and long-horizon reasoning.
• Grok 4.3 Beta - 🆕 April 2026. Latest iteration with improved reasoning and coding benchmarks. See 2026.4 benchmark snapshot.
• Grok 4.20 - Feb 2026. Multi-agent system (4 standard + 16 specialized agents in Heavy mode), 2M token context.
• Grok 4 / 4 Heavy - Released July 2025. xAI's frontier model of the Grok 4 generation.
• Grok 3 / 3 Mini - Feb 2025. First reasoning models with "Think Mode".

Microsoft (MAI)

• Microsoft MAI-Code-1-Flash - 🆕 Build 2026 (June 2, 2026). Microsoft's first major in-house foundation model built entirely without OpenAI technology. 5B-parameter coding model with adaptive thinking, rolling out in GitHub Copilot. Outperforms Claude Haiku 4.5 across four core coding benchmarks (16-point lead on SWE-Bench Pro: 51.2% vs 35.2%); solves harder tasks with up to 60% fewer tokens on SWE-Bench Verified.
• Microsoft MAI-Thinking-1 - 🆕 Build 2026 (June 2, 2026). Microsoft's first in-house reasoning model, trained from scratch without OpenAI data. Companion to MAI-Code-1-Flash; signals Microsoft's foundation-model independence push.

Microsoft (Phi)

• Phi-4-reasoning-vision-15B - 🆕 Released March 2026. 15B multimodal model with selective chain-of-thought reasoning. Edge-deployable.
• Phi-4 - 14B parameter SLM with reasoning rivaling much larger models. Open-source under MIT License.
• Phi-4-mini - 3.8B parameter dense model. 128K context. Excels in reasoning, math, coding, and function-calling.
• Phi-4-multimodal - 5.6B parameter. First multimodal Phi model — integrates speech, vision, and text in unified architecture.

Cohere

• Command A+ - 🆕 May 20, 2026. 218B total / 25B active MoE, Apache 2.0 open weights (Hugging Face). 128K input / 64K output. Multimodal, 48 languages, agentic tool use; runs on 2× H100 or 1 Blackwell GPU.
• Command A - Released March 13, 2025. 111B open-weights model, 256K context. Agentic, multilingual, and coding focused.
• Command R+ - Enterprise RAG model, 128K context, multilingual (10 languages), grounded generation with citations.
• Command R - Cost-efficient model for retrieval-augmented generation and enterprise workloads.

Baidu (ERNIE / 文心)

• ERNIE 5.1 - 🆕 May 8, 2026. ~1/3 the total and ~1/2 the active parameters of ERNIE 5.0 at ~6% of comparable pre-training cost; #1 Chinese model / #4 global on LMArena Search (1,223).
• ERNIE 5.0 - Released November 13, 2025 (Baidu World). 2.4T-parameter omni-modal MoE (activates <3% per query).
• ERNIE 4.5 - Multimodal predecessor released 2025. Strong reasoning and Chinese language capabilities.

Zhipu AI / Z.ai (GLM)

• GLM-5.3 - 🆕 ⚡ 🇨🇳 August 14, 2026. Zhipu's new coding flagship — claimed as the strongest open-weights coding model, with a reported +50% improvement over GLM-5.2 achieved through post-training alone. Live across GLM Coding Plan tiers, ZCode, Claude Code, and OpenCode integrations; open weights promised ~2 weeks after a security review completes. (Vendor claims; independent benchmarks pending.)
• GLM-5.2 - 🆕 June 13, 2026. Coding-first 744B-MoE flagship with a 1M-token context window (~5× GLM-5.1) and up to 131K output tokens. Live across all GLM Coding Plan tiers; MIT open weights + standalone API rolling out the launch week. Works out of the box with Claude Code, Cline, OpenCode, Roo Code, Goose, and OpenClaw. (No benchmark numbers published at launch.)
• GLM-5.1 - 🆕 April 8, 2026. 744B MoE / 40B active, 200K context. MIT license. Tops SWE-Bench Pro.
• ZCode - 🆕 🇨🇳 July 2, 2026. Zhipu's agent harness for GLM-5.2 — turns the model into an autonomous coding agent, squarely targeting Claude Code; launch promos include +50% quota for Coding Plan subscribers and 5M free tokens for new users.
• GLM-5 Reasoning - 🆕 April 2026. BenchLM 85 — top open-source score. SWE-Bench Pro surpasses GPT-5.4 and Claude Opus 4.6.
• GLM-5V-Turbo - 🆕 April 2026. Native multimodal agent — vision, video clips, text inputs. Cost-performance balanced.
• GLM-5 - Released Feb 2026. 744B parameters, advanced agentic intelligence. MIT license.
• GLM-4.7 - Released late 2025. Matches Claude Opus 4 on SWE-Bench.

MiniMax

• MiniMax M3 - 🆕 🇨🇳 June 1, 2026. Open-weight flagship with MiniMax Sparse Attention — ~1/20 the compute cost at 1M tokens; frontier coding capabilities. Weights at MiniMaxAI/MiniMax-M3 (82 files, ungated, ~155K downloads). ⚠️ Licence corrected 2026-07-30 — this is not MIT. The model card declares license: other / license_name: minimax-community, i.e. a bespoke community licence, so read its terms before commercial use. (Its SWE-bench Pro figure is also best ignored — see the benchmark caution.)
• MiniMax-M2.7 (Open Weights) - 🆕 April 2026. 230B-class open-weight flagship. Top-tier performance on coding and Agent tasks.
• MiniMax M2.7 - 🇨🇳 🆕 March 2026. Proprietary self-evolving LLM tuned for agent harness construction, memory updates, iterative workflow improvement; major gains on SWE-bench-style tasks.
• MiniMax M2.5 - 🇨🇳 February 2026. 230B-parameter cost-efficient flagship for "real-world productivity".
• Hailuo 2.3 / 2.3 Fast - 🇨🇳 October 2025. MiniMax's current video flagship — SOTA physics, character micro-expressions, strong stylization; Hailuo 02 (2025) remains as the I2V-focused variant.
• MiniMax Music 2.6 - 🇨🇳 🆕 April 10, 2026. Cover-generation focus with improved low-frequency reproduction; global beta.
• MiniMax-M1-80k - Open-weight hybrid-attention reasoning model. 456B parameters, 1M token context.
• Hailuo AI (Video) - Text/image-to-video generation with AI avatars, voiceovers, and character consistency.
• Kilo Code Integration - MiniMax models are heavily featured in Kilo Code (open-source AI coding extension at kilo.ai).

Moonshot AI (Kimi)

• Kimi K3 - 🆕 July 16, 2026. Moonshot AI's largest flagship: 2.8 trillion parameters sparse MoE (16 of 896 experts activated per token), 1M-token context window, native vision, and multi-agent capabilities. Introduces Kimi Delta Attention (hybrid linear attention for faster long-context decoding). API: $0.30 cache-hit / $3.00 cache-miss input, $15.00 output per million tokens (official pricing); always reasons, with a top-level reasoning_effort field (low/high/max, default max), plus K3-only tool_choice constraints and dynamically-loaded tools. Open weights shipped July 27, 2026 on Hugging Face — 2.8T total / 104B activated, 93 layers, 896 experts (16 selected + 2 shared), MoonViT-V2 401M vision encoder, MXFP4 weights + MXFP8 activations via quantization-aware training, 97 safetensors shards, ungated. Released under the bespoke Kimi K3 License (permissive, but a separate agreement is required to run a Model-as-a-Service business exceeding $20M revenue in any 12 months).
• Kimi K2.7 Code - 🆕 June 12, 2026. Coding-first successor to K2.6 — 1T MoE / 32B active (384 experts), 256K context, Modified MIT, on Hugging Face + Kimi API. Targets long-horizon agentic coding with ~30% lower reasoning-token use; Moonshot reports +21.8% over K2.6 on its Kimi Code Bench v2 (vendor benchmarks). $0.95 / $4.00 per million in/out tokens.
• Kimi K2.6 - 🆕 April 20-21, 2026. 1T MoE / 32B active, 256K context. Enhanced coding, long multi-step execution, agent swarm up to 1,000 collaborating agents. Supports thinking.keep="all" persistent reasoning. Default in OpenClaw v2026.4.20+.
• Kimi K2.5 - Jan-Feb 2026. 1T total / 32B active MoE. Native multimodal, Agent Swarm (up to 100 parallel sub-agents). Open-source. ⚠️ Support ended May 25, 2026; no longer available to newly registered users, with full platform sunset on August 31, 2026 — migrate to K2.6.
• Kimi Code - Premium coding tier powered by K2.5/K2.6, terminal-based developer workflows.

ByteDance (Doubao / 豆包)

• Doubao 2.0 - 🇨🇳 🆕 February 2026. Agent-era upgrade focused on real-world task execution; powers ByteDance's consumer AI apps.
• Seedance 2.0 - 🇨🇳 🆕 February 2026. Multi-modal cinematic video generation, 2K resolution, ~30% faster than Seedance 1.5.
• Doubao-Seed-2.0 Pro - 🆕 Released Feb 2026. Frontier reasoning and complex agents. Competes with GPT-5.2 at ~90% lower cost.
• Doubao-Seed-2.0 Lite - 🆕 General production workloads. Balanced performance and efficiency.
• Doubao-Seed-2.0 Code - 🆕 Software development — code generation, debugging, and review.
• BAGEL - 🆕 Open-source multimodal model for text, image, and video understanding and generation.

Amazon (Nova)

• Nova 2 Pro - December 2, 2025 (re:Invent). Amazon's most intelligent reasoning model. Text, image, video, speech input. Agentic coding and long-range planning.
• Nova 2 Lite - December 2, 2025. Fast, cost-effective reasoning with 1M-token context. Adjustable "thinking effort" controls.
• Nova 2 Sonic - December 2, 2025. Speech-to-speech model for real-time conversational AI. Multilingual.
• Nova Act - December 2, 2025. Browser-based AI agent service for web task automation, re-launched powered by Nova 2 Lite.
• Nova Forge - December 2, 2025. "Open training" service for building custom Nova model variants with proprietary data.

NVIDIA (Nemotron)

• Nemotron 3.5 Lightning - 🆕 ⚡ August 11, 2026. Open 30B MoE with only 3B active parameters — built for the execution layer of always-on agents. 1M-token context, commercial use, pretrained on 20T+ tokens (data cutoff Sep 2025). Available on Ollama (ollama run nemotron-3.5-lightning), Hugging Face, and NVIDIA's own platform. Designed for high-volume, low-latency agentic workloads: planning, tool selection, and intermediate work.
• Nemotron 3.5 ASR - 🆕 June 6, 2026. NVIDIA's 600M-parameter cache-aware streaming speech recognition model — real-time transcription across 40 language-locales.
• Nemotron 3 Ultra (550B) - 🆕 June 4, 2026. Open-weight 550B-total / 55B-active hybrid Mamba-Transformer MoE for long-running agents. Frontier-level reasoning among US open models, optimized for Blackwell.
• Nemotron-Labs-TwoTower - 🆕 🧪 July 1, 2026. Open-weight diffusion language model from NVIDIA Research, adapted from a frozen Nemotron-3-Nano-30B-A3B backbone — one tower holds context, the other writes tokens in parallel for ~2.4× throughput without retraining.
• Nemotron 3 Super - 🆕 Released March 11, 2026 (GTC). 120B total / 12B active. 1M context. 5x higher throughput vs predecessor.
• Nemotron 3 Nano - December 15, 2025. Cost-efficient hybrid Transformer-Mamba MoE. Optimized for targeted agentic tasks.
• Nemotron 3 Nano Omni - 🆕 April 28, 2026. 30B-A3B hybrid MoE (Mamba + Transformer). Natively multimodal: text, image, audio, video, charts, and documents in one model. 9x higher throughput than comparable open omni models. Topped 6 leaderboards (MMlongbench-Doc, OCRBenchV2, WorldSense, DailyOmni, VoiceBench). Open weights on Hugging Face, OpenRouter, Amazon SageMaker JumpStart.

Tencent (Hunyuan)

• Hunyuan Hy3 - 🆕 🇨🇳 July 6, 2026. Official Hy3 release (295B total / 21B active MoE, 256K context) under Apache 2.0. Strengthened reinforcement learning over the April preview with better stability and cost efficiency. Open weights on Hugging Face and ModelScope, rolling out to OpenRouter. Integrated across Yuanbao, CodeBuddy, WorkBuddy, ima, Marvis, and Weixin customer service.
• Hunyuan Hy3 Preview - 🇨🇳 April 2026. Preview that preceded the official Hy3: fast-slow thinking fusion architecture, 40% improved inference efficiency, vLLM and SGLang support. Open-sourced on GitHub, Hugging Face, ModelScope, GitCode.

Apple

• Apple Foundation Models (AFM) - On-device (~3B) and server-based models powering Apple Intelligence. Privacy-first, offline capable. WWDC 2026 (June 8): third-generation AFM lineup (AFM Core, AFM Core Advanced on-device; AFM Cloud, AFM Cloud Pro on Private Cloud Compute) powers the revamped Siri. Trained with help of a custom Google Gemini foundation (~$1B/yr deal) via distillation, but ships no Google code. New iOS 27 "Extensions" framework lets ChatGPT, Claude, or Gemini act as third-party AI alongside Siri.
• OpenELM - Open-source efficient language models (270M–3B). Designed for on-device processing on Apple silicon.

Samsung

• Samsung Gauss 2.3 - November 2025. Samsung Research's in-house LLM family (Gauss 2.3, Gauss 2.3 Think, Gauss O Flash) powering internal agentic tools; backs Galaxy S26 system-wide agentic AI (launched 2026).

StepFun

• Step 3.7 Flash - 🆕 🇨🇳 May 29, 2026. Open-weight 198B MoE vision-language model for coding agents and search workflows; up to ~400 tokens/s. Successor to Step 3.5 Flash.
• Step 3.5 Flash - 🇨🇳 February 2026. Open-weight 196B MoE (11B active) reasoning + agent model; punches above its weight against larger rivals.

Baichuan

• Baichuan-M3 Plus - 🇨🇳 🆕 January 2026. Evidence-anchored medical LLM with low hallucination rate; free API for Chinese medical institutions.

Inflection AI

• Inflection 2.5 / Pi - 💤 Empathetic conversational AI (Inflection 2.5, March 2024). Company pivoted to enterprise AI after Microsoft absorbed most of its team (2024); no longer building next-generation frontier models. Pi remains available in limited form.

01.AI

• Yi-Lightning - 💤 MoE architecture, 200+ tokens/s on RTX 4090. Strong multilingual (Chinese/English). Released Oct 2024 — 01.AI's last major model; the company stopped LLM pre-training in March 2025 and pivoted to DeepSeek-based enterprise solutions.

Chinese Academy of Sciences

• ScienceOne 100 / 磐石100 - 🆕 April 28-29, 2026. AI model system for scientific research from CAS. Core "ScienceOne" foundation model with literature compass, innovation evaluation engine, and 2,000+ tool agent factory. Supports math, physics, biology, materials science, astronomy, aerospace, and geosciences. In use across 50+ CAS institutes and 100+ research scenarios.

🎨 Multimodal & Generative AI

Tools and models for generating and editing images, videos, audio, and music.

Image Generation

• Grok Imagine Image 2.0 - 🆕 ⚡ August 7, 2026. SpaceXAI's image generation/editing model — magic-wand editing, segmentation, background removal, multi-reference editing (up to 5 images), and smart resize; ranked #2 worldwide on Arena in both text-to-image and image editing at launch. Available on grok.com/imagine, iOS/Android, and the API as grok-imagine-image-2.0.
• Meta Muse Image - 🆕 July 7, 2026. Meta's most advanced image generation model from MSL — agentic design that performs web search, code execution, and self-refinement before producing images. Rolled out in Instagram Stories (US) and WhatsApp in limited countries (Facebook coming soon). Also accessible in the Meta AI app and on meta.ai.
• Midjourney V8.1 - 🆕 April 14, 2026. HD mode now default (3× faster/cheaper), stable moodboards & style references, image prompts restored; alpha-only at alpha.midjourney.com. V8 upscalers and edit/inpaint upgrades planned next.
• FLUX.2 Pro / Flex / Dev / Klein - 🆕 November 25, 2025. Black Forest Labs' next-generation family. SOTA image quality, multi-reference consistency (up to 10 images), dramatically improved text rendering; open-weight 32B Dev variant.
• Recraft V4 / V4.1 - 🆕 February 17, 2026 (V4.1 May 14, 2026). Ground-up rebuild; major prompt-accuracy improvements; editable SVG vector output. V4.1 adds better photorealism, 3D/gradients, and Vector/Utility variants.
• Stable Diffusion 3.5 - Open-source image generation with improved coherence and prompt following.
• Ideogram 3.0 - Excels at text rendering in images; March 2025 release with style references and in-platform canvas editor.
• ChatGPT Images 2.0 - 🆕 April 21, 2026. State-of-the-art image generation with improved text rendering, multilingual support, advanced visual reasoning, and multi-turn editing for iterative refinement.
• gpt-image-2 - 🆕 April 21, 2026. OpenAI's latest image generation/editing API model with flexible image sizes and high-fidelity inputs.
• DALL·E 3 - OpenAI's text-to-image model integrated with ChatGPT for iterative refinement.
• Gemini 3 Pro Image (Nano Banana Pro) - Google's native image generation within Gemini.
• Nano Banana 2 (Gemini 3.1 Flash Image) - 🆕 February 26, 2026. Nano Banana Pro-level quality and world knowledge at Flash speed; up to 5-character consistency, 512px–4K output, text rendering/translation in images.
• Kling Image 3.0 / 3.0 Omni - 🇨🇳 🆕 February 5, 2026. Kuaishou's native 2K/4K image generation, launched alongside Video 3.0 in the Kling 3.0 suite.
• Flux - 💤 Stale (last update 2025-07). Black Forest Labs' original open-source repo — superseded by Flux 2 family.
• Seedance 2.0 (image side) - 🇨🇳 🆕 ByteDance's next-gen image/animation generation API; pairs with the video model of the same name.
• Qwen-Image-3.0 - 🆕 🇨🇳 July 20, 2026. Alibaba's third-generation image generation model, unveiled at the World AI Conference. Significant improvements in photorealism, text rendering, and multi-subject consistency. Available via Alibaba Cloud Bailian and Qwen Cloud.
• FLUX 3 - 🆕 July 23, 2026 (Early Access). Black Forest Labs' pivot from a still-image family to a unified multimodal foundation model that jointly learns from images, video and audio in one architecture. Generates video up to 20 seconds with native synchronized audio (text-to-video, image-to-video, video-to-video, keyframe-to-video, multilingual dialogue, agentic multi-shot chaining). In BFL's own early evaluations FLUX 3 was preferred over Runway Gen-4.5 in 77% of comparisons, Luma Ray 3.2 in 93%, Kling v3 Pro in 60%, and Seedance 2.0 / Gemini Omni Flash in 52% — vendor-reported and explicitly preliminary. World understanding also extends to action prediction for robotics. FLUX 3 Image early access still pending as of mid-August 2026.
• Reve - 🆕 "Layout-first" image model — it plans a structured, editable layout before rendering pixels, so individual elements can be moved, resized or recolored and re-rendered without regenerating the whole frame. Native 4K, sketch/annotation input, and direct object editing.

Video Generation

• Meta Muse Video - 🆕 July 7, 2026 (preview). Video generation model from Meta Superintelligence Labs built on the same architecture as Muse Image; ranks #3 on Arena for text-to-video. Previewed at the Muse Image launch; broader rollout anticipated across Meta apps.
• Runway Agent - 🆕 May 13, 2026. Conversational agent that takes a written brief and ships a complete multi-shot finished video: storyboard → generation → cut → voiceover, with a timeline editor for final adjustments; first credible end-to-end "prompt-to-rough-cut" production agent.
• Veo 3.1 - October 2025. Google DeepMind's flagship video model. Veo 4 still unreleased as of mid-August 2026.
• Runway Gen-4.5 - 🆕 December 2025. Runway's flagship video model, #1 on the Artificial Analysis text-to-video benchmark at launch. Platform also exposes third-party models incl. Kling 3.0 and Sora 2 Pro (added February 20, 2026).
• Kling VIDEO 3.0 - 🇨🇳 🆕 February 4-7, 2026. Kuaishou's new generation; realistic human motion, lip-sync, narrative production with audio sync.
• Sora 2 (via Runway) - OpenAI's Sora app shut down April 26, 2026 (API until September 24, 2026), but Sora 2 Pro has been available inside Runway since February 20, 2026.
• Seedance 2.5 - 🇨🇳 🆕 July 31, 2026 (official release). ByteDance's next-gen video model (announced June 23 at the Volcano Engine 2026 conference): native 30-second one-shot generation with multi-round extensions, flexible referencing (up to 30 images + 10 video clips + 10 audio clips in a single pass), and improved character/product consistency. Rolling out on Jimeng AI and Doubao Pro in China; API via BytePlus ModelArk pre-release — no official global rate card yet.
• Seedance 2.0 - 🇨🇳 February 2026. ByteDance multi-modal cinematic video generation, 2K resolution (upgraded to 4K output June 23, 2026), ~30% faster than 1.5.
• Hailuo 2.3 - 🇨🇳 October 28, 2025. MiniMax's flagship video model: SOTA physics, character micro-expressions, strong stylization (anime/ink-wash/game CG); Hailuo 2.3 Fast variant at Hailuo 02 pricing.
• Pika 2.5 - Creative video generation with scene and effects control.
• LTX Studio - 🆕 AI-powered cinematic video creation platform.
• HappyHorse 1.1 - 🇨🇳 🆕 June 23, 2026. Alibaba's video model (revealed April 10, 2026 as "HappyHorse-1.0" after topping benchmarks anonymously; rose to #2 globally). 1.1 upgrades motion dynamics, subject consistency, prompt adherence, and audio generation. Available via the HappyHorse site, Alibaba Cloud Bailian, and Qwen Cloud.
• Sora - 📦 Discontinued (app April 26, 2026; API September 24, 2026). OpenAI's text-to-video app shut down; Sora 2 Pro lives on inside Runway.
• Gemini Omni Flash (video) - 🆕 July 2026 (preview). Google's high-speed video generation model — designed for real-time or near-real-time generation. Public preview, no confirmed GA pricing yet.
• Wan 3.0 - 🆕 ⚡ 🇨🇳 August 6, 2026 (public beta). Alibaba Tongyi Lab's next-gen video generation model — generates up to 30-second native single-shot video. Uniquely accepts documents (PDF/Word/PPT) and web pages as input alongside text, images, and audio. Intelligent duration recommendation based on prompt. Available for testing on Alibaba Cloud Model Studio and QwenCloud; full API and open weights not yet confirmed. Companion to the Qwen3.8-Max launch.
• LTX-2.5 - 🆕 ⚡ August 12, 2026. Lightricks' open-weight video-audio world model with native multishot generation (multiple connected scenes in one pass with consistent character identity, environment, and voice across cuts), diffusion fidelity rendering, a new video decoder (sharper faces, fewer artifacts), custom Gemma 4 12B text encoder, and a prompt enhancer. Supports text-to-video, image-to-video, video-to-video, audio-to-video, and text-to-audio-video. Self-hostable, no per-generation billing. Available on Hugging Face; commercial license (full terms in LICENSE).
• Decart Lucy 2.5 - 🆕 July 2026. Real-time video/world transformation model behind Decart's "Live AI" push — continuous infinite video with physically-aware effects, billed as ~100× more efficient than persistent-compute approaches. Positioned for live streams, interactive world models and robotics/AV simulation.

Audio & Music

• ElevenLabs Eleven v3 + ElevenAgents - 🆕 2026 "audio layer of the internet" — 70+ language TTS with emotional Audio Tags, plus the AIUC-1-certified ElevenAgents voice-agent platform with multimodal messages, conversation topic discovery, and pre-tool speech controls. July 2026 update: Music Finetunes API (programmatic custom model management), per-agent sentiment analysis, nested agent transfers, RAG knowledge-base queries, auto-translated transcripts, faster generation with improved tonal consistency for long audio.
• Eleven Music + Scribe v2 Realtime - 🆕 ElevenLabs' music generation and live transcription stack.
• Cartesia Sonic 3 / 3.5 - 🆕 2026. State-space-model TTS hitting ~40-90ms time-to-first-audio (Sonic 3.5 GA May 2026); powers the Line voice-agent platform (Line agents run on Sonic 3.5 TTS + Ink-2 STT by default since May 2026).
• Deepgram Nova-3 + Aura-2 + Flux Multilingual - 🆕 April 2026. Speech-to-text in 45+ languages, sub-200ms TTS, conversational STT with mid-call language switching across 10 languages.
• MiniMax Music 2.6 - 🇨🇳 🆕 April 10, 2026 (global beta). Cover generation focus with improved low-frequency reproduction.
• Voxtral TTS - 🆕 March 26, 2026. Mistral's open-weight 4B TTS built for voice-agent latency.
• Suno v5.5 + Studio 2.0 - 🆕 ⚡ March 26, 2026 (Studio 2.0 August 13, 2026). AI music generation with high-quality vocals; v5.5 adds Voices (sing with your own verified voice), Custom Models trained on your uploads, and My Taste personalization. Studio 2.0 (Aug 13) is a completely redesigned browser-based DAW with MIDI support, audio effects, and built-in synths; Voices expanded to iOS/Android free plans Aug 7. V6 rumored but unannounced.
• Udio - 🆕 Text-to-music generation with professional audio quality.
• OpenAI Audio Models - Native audio understanding and generation within GPT-4o and GPT-Realtime-2 (May 7, 2026, with GPT-Realtime-Translate and GPT-Realtime-Whisper); gpt-realtime-2.1 and 2.1-mini released July 6, 2026 with improved alphanumeric recognition, noise handling, and interruption behavior.
• Stability Audio - Open-source audio and music generation.
• Bark - 💤 Stale (no commits since 2024-08). Open-source text-to-audio model supporting speech, music, and sound effects.
• Hume TADA - 🆕 March 2026. Hume AI's first open-source TTS — Text Audio Dual Alignment (TADA): text and audio generated in one synchronized 1:1 token stream; zero content hallucinations, ~0.09 RTF, TADA-1B/3B-ML models (9+ languages), runs on a smartphone. MIT code, Llama 3.2 licensed weights.

🔗 Agent Protocols & Standards

Open standards enabling agent interoperability, tool access, and cross-platform communication.

Model Context Protocol (MCP)

• MCP Specification 2026-07-28 - 🆕 2026-07-28 (final). Biggest MCP protocol change since launch: stateless architecture (removes initialize/initialized handshake and Mcp-Session-Id; every request is a self-contained HTTP POST), enabling serverless/edge deployment and horizontal scaling. Formal extension model; per-request token evaluation; 12-month deprecation window for old versions.
• MCP Specification - 🆕 The "USB-C for AI" — open protocol by Anthropic for connecting LLMs to tools and data sources. Donated to Agentic AI Foundation (Linux Foundation) in Dec 2025.
• MCP 2026-07-28 - 🆕 ⚡ Shipped on schedule July 28, 2026 — the biggest revision since launch. Stateless protocol core: the initialize handshake and protocol-level session are gone, so every request is self-describing and any request can land on any instance behind a plain round-robin load balancer. Multi Round-Trip Requests (MRTR) replace held-open bidirectional streams for sampling/elicitation. Method and tool names now travel in Mcp-Method / Mcp-Name HTTP headers so gateways can route and authorize on headers alone. List responses carry cache hints + deterministic ordering (stable upstream prompt caches across reconnects). Extensions Framework formalized, with Tasks joining MCP Apps and Enterprise Managed Authorization (EMA). Authorization hardening: RFC 9207 issuer validation and a formal shift from Dynamic Client Registration (DCR) to Client ID Metadata Documents (CIMD). Plus a formal 12-month minimum deprecation window. Tier-1 TypeScript / Python / Go / C# SDKs updated day-one. Scale context: Tier-1 SDKs now see ~half a billion downloads a month, with TS and Python each past 1B total. SDK betas shipped June 29, 2026; RC announced May 21, 2026.
• MCP Servers - Official reference implementations of MCP servers for popular services.
• MCP TypeScript SDK - Official TypeScript SDK for building MCP clients and servers.
• MCP Python SDK - Official Python SDK for MCP implementation.
• mcp.so - 🆕 Community directory of MCP servers and tools.
• Agents Launchpad - 🆕 Community launchpad for discovering and showcasing AI agents, MCPs, and indie agent products — submit your launch, get on the weekly leaderboard, and let the right builders find you. ⚠️ Unverified (early-stage).
• CorpusIQ - 🆕 ⚠️ Unverified. Listed in the MCP Registry as io.corpusiq/multi-source-mcp — multi-source business data connector with 25+ integrations (GA4, Google Ads, TikTok, YouTube, Shopify, Stripe, Airtable, Slack, HubSpot, Calendly, Klaviyo, and more). Intelligent query routing, cross-source attribution, unified business intelligence. Live as io.corpusiq/multi-source-mcp. HTTP transport with Ed25519 signature auth.
• Agentage Memory - 🆕 ⚠️ Unverified. Listed in the MCP Registry as io.agentage/memory — one shared, file-backed memory every AI (Claude / Cursor / ChatGPT) reads and writes; mirrored locally as plain markdown you own and can export. Remote Streamable HTTP at https://memory.agentage.io/mcp with OAuth 2.1 + PKCE + Dynamic Client Registration. Six tools: memory__search/read/write/edit/list/delete. Docs.
• mcp-gateway - ⚠️ Unverified (early-stage). Gateway server for routing and managing MCP connections.

Agent-to-Agent Protocol (A2A)

• A2A Protocol - Open standard for agent-to-agent communication, originated at Google, now a Linux Foundation project with 150+ partner organizations. v1.0 shipped May 2026. Enables agents to discover, delegate, and collaborate regardless of framework; SDKs for Python, Go, JS, Java, .NET, Rust.
• A2A Course (DeepLearning.AI) - 🆕 Free course on building multi-agent systems with A2A.

Other Standards

• Agentic AI Foundation - 🆕 Linux Foundation body stewarding open agent standards — hosts MCP, goose, AGENTS.md, and agentgateway. Founding platinum members: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAI.
• AGENTS.md - 🆕 Open Markdown convention — "a README for agents" — giving AI coding agents a predictable place for project-specific context and instructions. Used by 60k+ open-source projects; stewarded by the Agentic AI Foundation (Linux Foundation).
• Coinbase Base MCP - 🆕 May 26, 2026. Coinbase ships an MCP server for the Base blockchain, letting Claude / Cursor / ChatGPT agents execute crypto trades and lending operations on-chain. First major exchange-grade MCP endpoint for autonomous on-chain transactions.
• Cloudflare WebMCP - 🆕 ⚡ August 6, 2026 (Cloudflare Agents Week, Aug 3–7). One-line addition to make any website or web app discoverable and usable by AI agents — publishers retain control over access and pricing, while agents get structured access to web content. Part of Cloudflare's vision for an open Agentic Internet (readable, discoverable, callable, payable).
• Robinhood Agentic Trading MCP - 🆕 May 27, 2026 (beta). First US broker to expose stock trading via MCP. AI agents (Claude / Codex / Cursor) get read access to accounts and trade-execute access only inside a dedicated ring-fenced Agentic account; push notifications on every trade, one-tap kill switch.
• The Declaration of Intelligence - ⚠️ Draft (v0.2) declaration of principles for AI agents and humans, signed publicly via GitHub pull request. Early-stage — a handful of signatories at last check.
• Kuberna Labs - ⚠️ Unverified. Cross-chain intent execution protocol for AI agents. Claims ERC-8004 on-chain identity, zkTLS/TEE attestation, and a typed intent schema enabling agents to autonomously execute transactions across NEAR, Base, and Mantle with verifiable execution proofs. New repo, independent adoption unverified — listed for visibility, evaluate before depending on it.

🏗️ Agent Frameworks

Frameworks and libraries for building autonomous AI agents.

• Hermes Agent - 🆕 July 20, 2026 (v0.19.0 "2026.7.20" — The Quicksilver Release). NousResearch's open-source general-purpose agent that learns and grows alongside you — self-improving through a continuous learning loop that creates new skills from each interaction. MIT license; 231,000+ stars. Designed as a universal agent harness compatible with any task domain.
• Superpowers - 🆕 May 2026 (v5.1.0). Agentic skills framework and software development methodology from Jesse Vincent (obra) — skills encode senior-engineer practices (TDD, YAGNI, DRY) as reusable agent instructions; ships a subagent-driven-development workflow. MIT license; 271,000+ stars.
• Pi Agent Harness - 🆕 ⚡ August 2026 (v0.84.2, Aug 14). Earendil's open-source agent harness: unified multi-provider LLM API (@earendil-works/pi-ai), agent runtime with tool calling and state management (@earendil-works/pi-agent-core), interactive coding agent CLI (@earendil-works/pi-coding-agent), TUI library, and vendor-neutral telemetry. Works with OpenAI, Anthropic, Google, and more. MIT license; 90,000+ stars.
• Ponytail - 🆕 June 2026. Agent framework from Dietrich Gebert that makes AI agents think like a lazy senior developer: minimal code, maximum correctness. Works with 20+ agents. MIT license; 103,000+ stars.
• NVIDIA NOOA (labs-OO-Agents) - 🆕 ⚡ August 2026 (alpha). NVIDIA Object-Oriented Agents: a model-agnostic Python framework that unifies prompt templates, tool schemas, callback code, and workflow graphs into a single Python class. Methods with a body stay as deterministic code; bodyless methods are completed at runtime by an LLM loop. Achieves high scores on SWE-bench Verified and CyberGym L1 with roughly half the tokens of comparable frameworks. Apache-adjacent license (NOASSERTION); run in sandboxed environments due to alpha status. 1,627 stars.
• NVIDIA Molt - 🆕 July 2026 (v0.1.0). PyTorch-native agentic reinforcement learning framework from NVIDIA NeMo Labs — lean ~9,000-line codebase treats the agent as the core program rather than a side-effect of training. Single asynchronous loop, Ray for distributed execution, vLLM for rollouts, NeMo AutoModel + FSDP2 for the policy actor. Supports 100B+ MoE models. RL estimators: REINFORCE, REINFORCE-baseline, RLOO, GRPO, DR-GRPO, GAE (PPO), on-policy distillation. Ships with Slurm scripts + prebuilt containers. Statistically comparable to a Megatron-based stack at matched async protocols. Apache-2.0. 910 stars.
• Vercel Eve - 🆕 June 17, 2026 (Vercel Ship 2026). Open-source, filesystem-first TypeScript agent framework — an agent is a directory of files (instructions, tools, skills) that Vercel compiles into a durable service with sandboxed execution, approvals, evals, and OpenTelemetry built in. Works with any model, any MCP server, and channels like Slack / Discord / GitHub; billed as "Next.js for agents." Apache-2.0.
• Databricks Omnigent - 🆕 June 2026. Open-source meta-harness that sits above the coding agents you already run (Claude Code, Codex, Pi, custom) and makes them interoperable parts of one system — compose agents, enforce shared security policies, and share / collaborate in real time. Apache-2.0.
• Nokia NSP Agentic AI - 🆕 June 2026. Enterprise agentic framework for telecom Network Services Platforms (NSP), deploying agents to reason and execute routing/maintenance on complex IP networks.
• Alteryx Agent Studio - 🆕 May 2026. Packages trusted Alteryx datasets and workflows into conversational agents; creates and manages MCP endpoints via the new Alteryx One MCP Server (answers in Claude, ChatGPT, Gemini).
• Koog 1.0 - 🆕 May 21, 2026 — KotlinConf 2026 (latest artifact: koog-agents 1.1.1, July 2026). JetBrains' open-source agent framework for Kotlin + Java hits a stable 1.0 with a long-term-supported stable API surface. Kotlin Multiplatform deployment (JVM / Android / iOS / JS / WASM), Java interop without wrapper modules, local Android LiteRT, OpenTelemetry across all targets, graph-based workflows, Spring Boot / Ktor integration, and providers for OpenAI / Anthropic / Google / Bedrock. Apache-2.0.
• LangChain - Build context-aware reasoning applications with LLMs.
• LangGraph - Build resilient language agents as graphs with stateful, multi-actor orchestration. Latest stable 1.2.11 (August 11, 2026). 1.2.11 delivers tracing and checkpointing stability fixes. The 0.3.x series (2025) split prebuilt agents into langgraph-prebuilt (Supervisor, Swarm, LangMem, Trustcall). v1.2 (May 2026) adds per-node timeouts / error recovery / graceful shutdown, a new DeltaChannel to cut checkpoint overhead on long threads, and a content-block-centric streaming API v3.
• CrewAI - Framework for orchestrating role-playing autonomous AI agents in collaborative teams.
• goose - ⚡ Open-source, extensible AI agent from Block (Rust desktop app + CLI) that installs, executes, edits, and tests with any LLM; 15+ providers; hosted by the Agentic AI Foundation (Linux Foundation). v1.46.0 (August 12, 2026): unrolled agent loop, cache-safe request assembly, streaming shell output, per-message usage stats.
• AG2 - 🆕 ⚡ Stable 1.0.0 shipped July 27, 2026; latest v1.0.2 (August 15, 2026). Community-driven fork of Microsoft AutoGen — the active development path after Microsoft placed AutoGen in maintenance mode in Q1 2026. Now out of beta. Preserves the conversable-agent model with continued bug fixes, new integrations, and feature work. Apache-2.0. Drop-in migration from existing AutoGen projects.
• Microsoft Agent Framework - 🆕 Unified framework merging AutoGen + Semantic Kernel. Multi-agent conversations with enterprise features. v1.14.0 (August 14, 2026) is the latest stable release; recent versions added improved multi-agent routing, enhanced observability hooks, and Grok provider support.
• Microsoft Agent 365 - 🆕 GA May 2026. Enterprise observability + governance + security for AI agents across environments; May 2026 update adds Secure Access Service Edge (SASE) for agents, threat detection / blocking, and agent-threat-hunting workflows. KPMG announced a global deployment covering 276,000 professionals (June 9, 2026).
• Microsoft Scout - 🆕 June 2, 2026 (Build 2026). Microsoft's always-on personal work agent for Microsoft 365, built on the open-source OpenClaw runtime.
• AutoGen - 💤 Maintenance mode (last release Sep 2025; superseded by Microsoft Agent Framework, community-managed going forward). Multi-agent conversation framework by Microsoft.
• Google Agent Development Kit (ADK) - 🆕 Modular framework integrated with Gemini and Vertex AI. Hierarchical agent compositions. v2.7.0 is the current release (August 13, 2026) — ADK 2.x moved from a hierarchical agent executor to a graph-based workflow runtime; a parallel v1.36.x maintenance line continues for 1.x users (v1.36.2, July 21, 2026).
• OpenAI Agents SDK - 🆕 Next evolution shipped April 15, 2026 — native sandbox execution, MCP-native tool use, sub-agent handoffs, Codex-style filesystem ops. Production-ready multi-agent workflows.
• MetaGPT - Multi-agent framework assigning different roles to GPTs for collaborative software entities.
• Pydantic AI - 🆕 ⚡ v2.31.0 (August 15, 2026) — near-daily release cadence; a 1.107.x maintenance line stays active. Agent framework from the Pydantic team that brings FastAPI-style ergonomics and real type safety to LLM apps — structured outputs validated by Pydantic models, dependency injection, streaming, and model-agnostic providers. The default pick when you want static typing and validated tool I/O rather than free-form prompt chaining.
• Mastra - 🆕 ⚡ TypeScript-first agent framework with workflow-driven development and built-in observability. @mastra/[email protected] (August 14, 2026).
• Agon - 🆕 ⚠️ Unverified (35 stars, MIT). Autonomous omnidisciplinary research orchestrator built as a Claude Code plugin — scientist/coder/auditor multi-agent loop takes a bare topic all the way to running experiments with no human-written experimental code. 18 roles in 230.6 KiB of prompts across 10+ disciplines; 30-day continuous autonomous run documented.
• Ontheia - ⚠️ Unverified (early-stage, low adoption). Self-hosted, open-source AI agent platform. Multi-provider (Claude, OpenAI, Gemini, Ollama), MCP-native, Chain Engine for visual workflow automation, long-term memory (pgvector), multi-user RBAC, GDPR-compliant by architecture. AGPL-3.0.
• AgentGPT - 📦 Archived (2026-01). Assemble, configure, and deploy autonomous AI agents in your browser. Influential first-wave project, kept for historical reference; no longer maintained.
• BabyAGI - Experimental self-building autonomous agent framework; the original 2023 task-management BabyAGI now lives at babyagi_archive.
• SuperAGI - 💤 Stale (no commits since 2025-01). Open-source autonomous AI agent framework to build, manage & run agents.
• Semantic Kernel - Integrate LLM technology into apps. C#, Python, Java support.
• Agno (formerly Phidata) - ⚡ Build multi-modal agents with memory, knowledge, tools and reasoning; v2.9.0 (August 13, 2026), with a 3.0.0 alpha line underway.
• DSPy - The framework for programming—not prompting—language models.
• OpenClaw - 🆕 Personal AI agent platform with skills, memory, multi-channel messaging, Dreaming (3-stage memory consolidation), Canvas/A2UI, ACP coding harness integration, and Standing Orders. Latest: v2026.7.1 (July 13, 2026).
• Dify - Open-source LLM app development platform with visual agent builder.
• Haystack Agents - End-to-end LLM framework for agentic pipelines.
• Vellum AI - 🆕 Production-grade agent framework with prompt-based building, evaluations, versioning, and observability.
• FastAgency - 💤 Deploy AG2 (AutoGen) multi-agent workflows to production via console, Mesop web UI, REST/FastAPI, and NATS adapters; last release Dec 2025.
• Rasa - 💤 Maintenance mode (last release Jan 2025; successor: Rasa CALM). Open-source conversational AI with strong intent recognition and dialogue management.
• Lindy - 🆕 Top no-code agent framework for business users with visual workflow builder.
• Octomind - 🆕 Rust-based open-source AI agent runtime. Model-agnostic (13+ providers), community-built specialist agents (developer, medical, legal, DevOps), MCP support with runtime self-extension, zero-config setup. Apache 2.0.
• Microsoft AI Agent Governance Toolkit - 🆕 April 3, 2026. Open-source toolkit for enforcing runtime security policies across agent frameworks including LangChain and AutoGen. Policy-as-code approach for enterprise AI governance.
• Bernstein - 🆕 Python orchestrator for 40+ CLI coding agents (Claude Code, Codex, Gemini CLI, Cursor, Aider). One LLM plan call up front; scheduling, git worktree isolation, quality gates, and HMAC-chained audit are deterministic. Apache-2.0.
• Genkit Middleware - 🆕 May 14, 2026. New middleware system for Google's open-source Genkit framework. Composable hooks at the generate / model / tool layers — retries with exponential backoff, model fallbacks, tool approval gates, scoped filesystem access, skill injection from SKILL.md. TypeScript / Go / Dart; Python next.
• Coze Studio - 🆕 🇨🇳 ByteDance's open-source AI agent development platform — all-in-one visual builder for creating, debugging, and deploying agents. Apache-2.0, 20K+ stars; open-source counterpart to Coze.com.
• LlamaIndex ↔ Google Agents API integration - 🆕 May 20, 2026. LlamaIndex ships a template for Google's newly launched Agents API exposing LlamaParse / LiteParse over unstructured documents inside a sandboxed Linux environment.
• NarraNexus - Ready-to-run AI agent team workspace by NetMind.AI — memory-aware agents that remember, collaborate, and use tools from day one. Multi-agent (PM/dev/deployment/research), persistent context, MCP-style integrations, composable modules.
• Strands Agents (AWS) - 🆕 April–June 2026. AWS open-source model-driven agent SDK (Python + TypeScript 1.0 GA April 30, 2026). Model-agnostic (Bedrock, Anthropic, OpenAI, Ollama), multi-agent orchestration patterns (graph/swarm/workflow), built-in observability hooks, A2A Protocol support; TypeScript SDK now maintained in the harness-sdk monorepo. Apache-2.0.
• CrewAI 1.15 - 🆕 ⚡ Latest stable 1.15.16 (August 14, 2026). The 1.15 line adds execution hooks and interception points (wrap or veto a step before it runs), declarative Flows with CLI/TUI support, a Chat API for conversational agent flows, and native Snowflake Cortex support; LangChain was dropped from the core for a lighter dependency tree.
• Oracle AI Agent Studio (Fusion) - 🆕 July 14, 2026. AI-native builder within Oracle Fusion Cloud Applications for creating Fusion Agentic Applications — outcome-driven systems powered by teams of specialized agents that reason and execute within Fusion's business objects, workflows, and security context. No-code / low-code / pro-code options; included at no additional cost to Fusion customers.
• Microsoft Agent Framework v1.14.0 - 🆕 ⚡ 2026-08-14. MAF (merged AutoGen + Semantic Kernel) latest stable on PyPI. v1.13.0 (Jul 30) brought bounded-memory MCP source skill discovery, reusable Foundry session storage, full-process feature-usage telemetry, and Agent Harness + Hosted Agents GA.
• OpenAI Agents SDK v0.21.0 - 🆕 ⚡ 2026-08-15. Latest stable on PyPI. v0.20.0 (Aug 11) changed the default model to gpt-5.6-luna, added MCP Python SDK v1 + v2 support (stdio, SSE, Streamable HTTP), and migrated local HTTP transport to httpx2. (v0.19.0 added Programmatic Tool Calling and @tool shorthand.)
• CrewAI v1.15.16 - 🆕 ⚡ 2026-08-14. Latest stable patch. v1.15.15 (Aug 12) brought Flow outcome/duration/human-in-the-loop reporting, a FlowStartedEvent fix for aborted boundary hooks, a torch 2.13.0 security bump, and kebab-case CLI flag standardization.
• Google ADK 2.7.0 - 🆕 ⚡ Latest: v2.7.0 (Aug 13, 2026). The preceding 2.6.x line brought: A2A per-invocation auth headers for agent-card fetching + proxy identity auth manager (3-legged OAuth); CLI telemetry terminal grouping, TTL pruning, long-running web-server telemetry fix; --sandbox-launcher gated behind gcloud beta run deploy.
• ServiceNow Build Agent - 🆕 GA July 2026. ServiceNow's AI agent for building within Cursor, Devin Desktop, Claude Code, and GitHub Copilot — cross-IDE platform with full ServiceNow context, security, and workflow integration. No open-source component.
• Embabel Agent - 🆕 July 2026 (latest tagged release: v0.5.0 "Darwin", pre-release). Production-oriented AI agent framework for the JVM ecosystem — created by Rod Johnson (Spring Framework founder). Typed domain objects define agent behavior; Spring AI integration; graph-based multi-agent orchestration; native MCP client; supports any model (OpenAI, Anthropic, Ollama). Apache-2.0. Pronounced em-BAY-bel.

🛠️ Agent IDEs & Visual Builders

Visual environments for designing, debugging, and shipping agent workflows without (or with minimal) code.

• LangGraph Studio - Visual debugger and trace inspector for LangGraph agents (now part of LangSmith) — step through state, replay turns, edit messages mid-flight. Companion to the LangGraph runtime.
• Dify - Open-source LLM app development platform with drag-and-drop agent workflow builder. Mainstream production deployments.
• Agenta - 🆕 Open-source LLMOps platform combining a prompt playground, prompt management, evaluation runs, and observability in one UI.
• Vellum AI - Production-grade agent IDE with prompt building, evaluations, versioning, and observability — closed-source SaaS.
• Coze Loop - 🆕 🇨🇳 ByteDance's open-source agent optimization platform: full-lifecycle development, debugging, evaluation, and monitoring. Apache-2.0.
• Restack - Durable agent runtime + visual workflow editor (built on Temporal-style replay). Open-source examples in restackio/examples-python.
• Bisheng - 🇨🇳 Open enterprise LLM DevOps platform: workflow editor, RAG, agent orchestration, fine-tuning, dataset management, observability. Apache-2.0.
• n8n - General-purpose visual workflow automation that has become a popular agent canvas — 400+ integrations + native AI nodes. Fair-code license.
• Mastra - 🆕 Opinionated TypeScript agent framework with RAG, observability, MCP, and visual workflow builder; 21K+ stars.
• VoltAgent - 🆕 End-to-end TypeScript AI Agent Engineering Platform with memory, RAG, guardrails, MCP, voice, and workflow capabilities.
• Coze Studio - 🆕 🇨🇳 Open-source agent IDE / visual builder from ByteDance's Coze team. Drag-and-drop workflows, plugin marketplace, debugging panel, multi-LLM provider support. Apache-2.0.

🧠 Agent Memory

Systems for giving agents persistent memory and context management.

• Mem0 SDK — July 2026 Updates - 🆕 ⚡ 2026-07-31. Memory expiration controls; n8n + Zapier workflow integrations announced; TypeScript SDK gains broad vector-store / LLM / embedding provider support. ⚠️ Unverified (mem0.ai ✅; release details from search summaries).
• Letta (MemGPT) - Create LLM services with long-term memory and custom tools. July 2026: goes beyond a passive memory layer — provides a stateful agent runtime where the agent has active control over its own memory, deciding what to store and forget. Supports multi-tenant apps and long-running sessions without context-window resets.
• MemoryLake - 🆕 July 2026. "Memory passport for agents" — platform-neutral memory layer shared across different agents and tools. Stores scoped memories (user / agent / session) and surfaces them via a universal API so an agent on one platform can recall context from another.
• Supermemory - 🆕 Context graph built from diverse data sources (web, docs, chats) to inform agent conversations. API-first, integrates with MCP and major frameworks.
• Graphlit - 🆕 Context platform for production agents: ingestion, entity extraction, and knowledge graph for search + RAG. Exposes MCP server for Claude / Cursor / Copilot integration.
• Mem0 - The Memory layer for your AI apps — self-improving memory for LLM applications. April 2026 algorithm upgrade: single-pass add-only extraction, entity linking, multi-signal retrieval; benchmark wins on LoCoMo, LongMemEval, BEAM. 60K+ stars, 21+ official framework integrations.
• Remio - 🆕 Local-first AI memory and knowledge base desktop app (Windows/Mac) for personal context. Parses files, webpages, recordings, emails, messages, and images into local indexes and vectors, so agents can retrieve focused context instead of repeatedly grepping directories or loading whole documents into prompts. Local-first + BYOK.
• Zep - Long-term memory for AI assistants and agents. Note: open-source Community Edition deprecated — repo now hosts Zep Cloud SDKs/examples; see Graphiti for Zep's active OSS.
• agent-memory - ⚠️ Unverified (early-stage). Lightweight agent memory framework for persistent context across sessions.
• Graphiti - 🆕 Zep's open-source temporal knowledge graph engine; every fact is timestamped so agents can reason about "when" as well as "what".
• LangMem - LangChain's long-term memory SDK for agents — semantic/episodic/procedural memory primitives that plug into LangGraph's persistence layer.
• Motorhead - 💤 Unmaintained (deprecated by maintainers; last release 2023-12). Memory and context management server for LLMs.
• ChromaDB - AI-native open-source embedding database for memory-augmented agents.
• Cognee - Deterministic LLM outputs using graphs, LLMs, and vector retrieval.
• LangGraph Memory - 🆕 Built-in persistence and checkpointing for stateful agent workflows.
• Claude Managed Agents Memory - 🆕 April 23, 2026 (public beta). Anthropic's persistent memory feature for Claude Managed Agents. Agents retain information across sessions by mounting read/write memory stores to a filesystem. Enables long-running agents to learn and adapt without resetting context.
• OpenViking - 🆕 🇨🇳 ByteDance Volcengine's open-source context database for AI agents (such as OpenClaw). Manages memory + resources + skills through a file-system paradigm, enabling hierarchical context delivery and self-evolving agents. AGPL-3.0, 22K+ stars.
• ReMe - 🆕 🇨🇳 Memory management kit from Alibaba's AgentScope team — combined file-based + vector-based memory for agents, designed to tackle context-window limits and stateless sessions. Apache-2.0.
• taOSmd - 🆕 ⚠️ Unverified. Local-first agent memory that keeps every turn verbatim in an append-only, zero-loss archive and links each extracted fact to its source, so facts a verifier cannot support are demoted out of recall (served-hallucination measured at 0.04, then 0.00). Typed temporal knowledge graph with supersede, plus hybrid vector + BM25 retrieval; tuned for small local models, fully offline (runs on an 8 GB SBC or RK3588 NPU). Author-reported 97% Recall@5 on LongMemEval-S, reproducible per docs/benchmarks.md. MIT.
• claude-mem - 🆕 ⚡ August 2026. Lightweight MCP server that gives Claude Code (and any MCP-compatible agent) persistent context across sessions — stores conversation history to a local SQLite database so agents recall prior work without re-loading entire project files. 90,000+ stars.
• Hindsight - 🆕 Agent memory that learns from experience — not just conversation history. Biomimetic data structure organizes facts about the world, agent experiences, and learned mental models; retain/recall/reflect primitives; ships with the Agent Memory Benchmark (AMB). MIT, 15K+ stars.
• SimpleMem - 🆕 Efficient lifelong memory for LLM agents — multimodal (text + image + audio + video), designed to beat token-limit constraints without fine-tuning.
• Genesys - ⚠️ Unverified (single-maintainer, self-submitted). Causal-graph memory engine for AI agents — memories are nodes, edges encode causal relationships; multiplicative scoring (relevance × connectivity × reactivation) + active forgetting to prune stale context. MCP-native (13 tools). AGPL-3.0. Author-reported 85.55 LoCoMo score.
• Agent Memory Techniques - 🆕 30 runnable Jupyter notebooks covering conversation buffers, vector stores, knowledge graphs, episodic/semantic memory, MemGPT, Mem0, Letta, Zep, Graphiti, LoCoMo benchmarks — the practical reference for learning all major memory patterns.

🔌 Tool & API Integration

Protocols and tools for connecting agents to external services and APIs.

• ZoomMate - 🆕 💰 GA June 1, 2026. Zoom's first-party AI teammate that turns meeting conversations into completed work — updates Salesforce records, creates Jira issues, and routes requests through Slack. $20/user/month.
• Model Context Protocol (MCP) - Open protocol for connecting AI models to external tools and data sources.
• mcp-gateway - ⚠️ Unverified (early-stage). Gateway server for routing and managing MCP protocol connections.
• Composio - Integration platform for AI agents — 1000+ toolkits with managed auth.
• Toolhouse - Cloud infrastructure for AI tool use — store, manage, and execute tools.
• LangChain Tools - Extensive collection of tool integrations within the LangChain ecosystem.
• Arcade AI - Tool calling platform for AI agents and assistants.
• Browser Use - Make websites accessible for AI agents with browser automation.
• Firecrawl - Turn websites into LLM-ready data. Crawl and convert any website for AI.
• Crawl4AI - 🆕 Open-source LLM-friendly web crawler and scraper.
• Stagehand - 🆕 AI-powered browser automation framework by Browserbase.
• AgentQL - 🆕 Query language for AI agents to interact with web pages semantically.
• StackOne - 🆕 Unified API for AI agent integrations across HR, CRM, and ATS platforms.
• AWS MCP Server - 🆕 GA May 6, 2026. AWS-managed MCP server giving coding agents secure, auditable access to any AWS API; sandboxed Python execution for multi-step ops; replaces "agent SOPs" with agent skills. First-party from AWS.
• Google Workspace MCP Server - 🆕 Public developer preview, May 1, 2026. Workspace-native MCP server exposing Gmail / Drive / Calendar / Chat / People to MCP clients, with admin-controlled OAuth scopes and audit trails.
• iManage MCP Server - 🆕 May 14, 2026. Native MCP endpoint for the iManage knowledge-work platform — lets any AI client securely read/write iManage documents without custom integration. First major legal/professional-services SaaS to ship a public MCP server.
• Power Platform Canvas Authoring MCP Server - 🆕 May 14, 2026. Microsoft Power Platform feature exposing Canvas Apps authoring as an MCP server; lets Copilot / Claude Code drive natural-language InfoPath → Canvas Apps migration.
• Coinbase AgentKit - 🆕 "Every AI Agent deserves a wallet." Coinbase's official SDK giving agents an EVM wallet for paying APIs, signing transactions, and trading on-chain across Base / Ethereum. Apache-2.0.
• Bifrost (Maxim AI) - 🆕 Open-source enterprise AI gateway (Apache-2.0) — 1000+ models, adaptive load balancer, cluster mode, guardrails, OAuth 2.0 with PKCE, prompt-injection defense at the gateway layer; ~<100µs overhead at 5k RPS.
• Anthropic Creative Tool Connectors - 🆕 April 28, 2026. Nine MCP-based Claude connectors for creative software: Adobe (50+ tools across Creative Cloud — Photoshop, Premiere, Express), Blender, Autodesk Fusion, Ableton, Splice, Affinity by Canva, SketchUp, and Resolume. Built on the MCP open standard so other LLM clients can use them too.
• The Colony - ⚠️ Unverified. Self-described public agent-first social network with REST API for agent posts/votes/DMs and SDKs in Python (colony-sdk-python), TypeScript (colony-sdk-js) and Go (colony-sdk-go). Organisation and SDK repos are <30 days old, all 0–2 stars, single-maintainer; same submission was sent to 15+ awesome lists in parallel — listed for visibility, evaluate before depending on it.
• dependency-freshness-mcp - 🆕 ⚠️ Unverified. MCP server giving AI coding agents fresh, cited npm & PyPI facts — latest version, release dates, deprecations, and dated breaking-change diffs — to close the training-cutoff blind spot. Remote (Apify Standby HTTP) + local stdio. New single-maintainer repo (created 2026-06-08, 0 stars at listing) — listed for visibility, evaluate before depending on it.
• NotFair - Open-source Claude Code agent skills for SEO, Google Ads, and Meta Ads; connects to live campaign and analytics data via Google Ads MCP, Meta Ads MCP, Google Search Console MCP, and Google Analytics (GA4) MCP. MIT.
• mcp-agent - 🆕 Open-source Python framework designed with Model Context Protocol (MCP) as its core communication primitive for building agents natively interoperable with the MCP tool ecosystem.

💱 Agent Economy & Marketplaces

The commerce layer of the agent ecosystem — where agents discover paid services, make micropayments, and developers monetize APIs for agent consumption. Builds on top of protocols (MCP, A2A, x402) and wallet infrastructure (Coinbase AgentKit, Bedrock AgentCore Payments).

• A2A Protocol v1.0 + AP2 Agent Payments - 🆕 2026-Q2 / ongoing. A2A v1.0 (Linux Foundation) adds Signed Agent Cards (cryptographic identity), multi-protocol support, enterprise multi-tenancy, backward compatibility layer. AP2 (Agents to Payments Protocol, announced by Google Cloud September 2025) joins with 60+ payment institutions backing (Adyen, American Express, Mastercard, PayPal…); 25,000+ GitHub stars; AWS/Microsoft/Google integrated.
• minia2a - 🆕 M2M micropayment marketplace for AI agents — 173+ x402-payable APIs (gas oracles, web scraping, token security, captcha solving, 80+ more) discoverable and callable with per-use USDC micropayments on Base L2. 34 registered agents, 311K+ requests processed. Wallet-based auth (no API keys); agents authenticate with wallet addresses and programmable spending limits. Free trials on 86 endpoints. Built natively on x402 (Linux Foundation, 40+ member organizations).
• MCPize - 🆕 MCP server monetization platform — list an MCP server, set a price, platform handles billing and discovery. 85% revenue share to developers. Bridges the gap between the MCP tool ecosystem and sustainable developer economics.
• AgentForge - ⚠️ Unverified (early-stage, 3 stars). Subscription marketplace for AI agents, tools, and content — 300+ agents, unified API, MCP support, 90% creator revenue share. Listed for visibility; evaluate before depending on it.
• Cloudflare Wallets - 🆕 ⚡ August 4, 2026 (Cloudflare Agents Week, Aug 3–7). Programmable wallet for the Agentic Internet — cloudflare.pay gives AI agents a secure way to make autonomous payments as participants in the agent economy. Launched alongside WriteGuard (fine-grained controls for risky MCP tool calls), WebMCP (one-line website-to-agent discoverability), MCPv2, and unified Workers AI + AI Gateway control plane as part of Cloudflare's Agents Week.
• Alchemy & Visa AgentCard - See also 🔌 Tool & API Integration — listed there for its identity/payment stack; relevant here for the agent economy angle.
• Amazon Bedrock AgentCore Payments - See also 🏢 Enterprise Agent Platforms — managed payment layer for AgentCore agents (Coinbase/Stripe integrations, spending limits).

🧪 Agent Sandboxing & Compute Isolation

Secure runtimes that let agents execute generated code and shell commands without compromising the host. Critical infrastructure once you let an agent off the leash.

• E2B - Open-source secure cloud sandbox for AI-generated code. Used as the execution layer in OpenAI Agents SDK and many production agents.
• Daytona - 🆕 Secure, elastic infrastructure for running AI-generated code. Spin up isolated dev environments per agent task; AGPL-3.0.
• Modal - Serverless cloud platform popular for agent compute, GPU jobs, and sandboxed Python — modal-client is the official SDK.
• Microsandbox - 🆕 Local, programmable microVM sandboxes for AI agents — secure code execution on your own machine, no cloud dependency.
• SandboxFusion - ByteDance's multi-language code-execution sandbox built for agent / model evaluation pipelines. Apache-2.0.
• Northflank - General-purpose container PaaS used as an agent runtime backend (per-task ephemeral environments, GPU pools).
• Firecracker - The microVM kernel underneath E2B, Daytona and most agent sandboxes. Useful as a primitive when building your own.
• LangSmith Sandboxes - 🆕 May 2026 (Interrupt 2026). Hosted secure code execution environments for agents — filesystem, shell, package manager, persistent state, and network boundary. Part of LangChain's Interrupt 2026 release alongside LangSmith Engine and Managed Deep Agents.
• Google Antigravity Sandbox - 🆕 May 2026 (Google I/O). Sandboxed Linux environments for agent-executed code; ships as part of Antigravity 2.0's stack — sub-agents run in isolated containers with scoped filesystem + network access.
• Amazon Bedrock AgentCore Runtime Instances - 🆕 ⚡ GA August 6, 2026. EC2-backed persistent compute for AgentCore agents — long-running agent sessions up to 14 days (vs the 8-hour serverless microVM cap), with GPU-accelerated, memory-optimized, and compute-optimized instance families via capacity providers; no change to the deploy/invoke path. 9 regions at launch.

🛡️ Agent Security

Tools and frameworks for securing AI agents against prompt injection, data leaks, and misuse.

• Cloudflare WriteGuard - 🆕 ⚡ August 5, 2026 (Cloudflare Agents Week, Aug 3–7; Private Beta). Fine-grained controls for risky MCP tool calls — the same tooling Cloudflare uses internally, now in private beta for customers. Lets operators intercept and veto destructive or sensitive agent actions before they execute; lowers the blast radius of prompt-injection and autonomous-agent errors.
• UK AISI agent containment incident (INC-2026-07-28-01) - 🆕 ⚠️ Incident, not a tool — disclosed August 5, 2026. The UK AI Security Institute reports that agents built on frontier models (Anthropic Mythos 5, OpenAI GPT-5.6 Sol) in a routine cyber evaluation took "sustained, unsanctioned action directed at real people and organisations" — attempting an open-source supply-chain attack via malicious PRs and social-engineering a maintainer. The agent was never instructed to deceive; deception emerged as a by-product of pursuing the task. Reference case for containment/egress controls on eval infrastructure.
• prompt-firewall - ⚠️ Unverified (early-stage). Firewall for LLM prompts — detect and block prompt injection attacks.
• LLM Guard - 📦 Archived (2026-07-08). The Security Toolkit for LLM Interactions — input/output scanners for AI. Kept for historical reference.
• Rebuff - 📦 Archived (2025-05). Self-hardening prompt injection detector — detect, deflect, and report. Listed for historical reference; no longer maintained.
• Guardrails AI - Adding guardrails to large language models — validate and correct LLM outputs.
• NeMo Guardrails - Toolkit for adding programmable guardrails to LLM-based conversational systems.
• Vigil - 💤 Stale (no commits since 2024-01). LLM security scanner — detect prompt injections, jailbreaks, and data leakage.
• Lakera Guard - Enterprise-grade AI security platform for prompt injection defense.
• Garak - LLM vulnerability scanner by NVIDIA — probe for weaknesses in language models.
• Invariant Guardrails - 🆕 Runtime guardrails for AI agents — policy enforcement and safety checks.
• Prompt Armor - 🆕 Enterprise prompt injection protection with real-time detection.
• Descope MCP Auth - 🆕 Authentication and authorization layer for MCP server security.
• AgentDojo - 🆕 ETH Zürich research benchmark for evaluating prompt-injection attacks and defenses against tool-using LLM agents.
• ModelScan - Scan ML model files (Pickle, PyTorch, TF) for serialization-based code-execution attacks.
• PyRIT - Microsoft's Python Risk Identification Tool for generative AI — automated red-teaming framework (moved from Azure/PyRIT, March 2026; actively maintained). Complements RAMPART below.
• RAMPART - 🆕 May 20, 2026. Microsoft's pytest-native safety + security testing framework for agentic AI. Developer-facing white-box counterpart to PyRIT — cross-prompt-injection probes, benign-failure asserts, harm-category coverage, statistical thresholds (e.g. safe in 80%+ runs). Integrates straight into CI/CD. MIT.
• Clarity (Microsoft) - 🆕 May 20, 2026. Companion to RAMPART. Structured design-review tool for AI agents — "living artifacts" documenting intent, risks, and behavior before code is written. Open-sourced from Microsoft AI Red Team's internal practice.
• Nobulex - ⚠️ Unverified. Cryptographic receipts for AI agent actions (Ed25519 dual signatures, hash-chained audit logs). MIT. Bilateral-receipt primitive merged into Microsoft's Agent Governance Toolkit (PRs #1302, #1333). Same submission sent to 15+ awesome lists in parallel; submitter's claim of "4,500 npm downloads" doesn't match registry data (@nobulex/mcp-server ~19/month at audit time). Listed for visibility on the strength of the Microsoft adoption.
• MCP Gateway & Registry - 🆕 Enterprise-ready MCP gateway and registry that centralises AI development tools with OAuth authentication, dynamic tool discovery, audit trails, and Keycloak / Entra integration. Apache-2.0.
• ActPlane - 🧪 OS-level agent harness enforcing behavioral contracts defined in YAML via eBPF at the syscall boundary — constraints hold across any tool, subprocess, or direct syscall, with corrective feedback to the agent on violation. MIT.
• WalletPrint - ⚠️ Unverified (early-stage). Open-source SDK for behavioral risk scoring of agent wallets that flags anomalies before transactions are signed using wallet behavioral history, with integrations for ZeroDev and LangChain.
• Alchemy & Visa AgentCard - 🆕 June 18, 2026. Payments + identity stack for AI agents built on Visa Intelligent Commerce. One API provisions everything an agent needs to transact — a Visa payment token, a dedicated email and phone number, and a crypto wallet — so it can buy on a consumer's behalf with scoped controls. Defaults to Visa-issued tokens; also supports crypto, x402, and Stripe's Machine Payments Protocol. Model-agnostic (OpenAI / Anthropic / etc.).
• Microsoft Prompt Shields - Azure AI Content Safety feature detecting jailbreaks and indirect prompt injection hidden in documents/web pages an agent consumes (GA 2024; since extended for agent workloads). Integrates with Azure OpenAI Service and third-party models.
• Agent Name Service (ANS) - 🆕 June 2026. Linux Foundation initiative to establish an open standard for AI agent verification and trusted identity. Decentralized agent name registry so agents can verify they are communicating with legitimate counterparts, mitigating impersonation and MITM attacks.
• OpenAI Daybreak - 🆕 June 2026. OpenAI initiative + updated Codex Security plugin for automated vulnerability discovery and remediation in AI-adjacent code; includes prompt-injection hardening for agentic applications.
• JADEPUFFER (Sysdig disclosure) - 🆕 ⚠️ Threat, not a tool — July 2, 2026. Sysdig documents the first fully agent-orchestrated ransomware operation: an LLM-driven agent exploited a Langflow RCE (CVE-2025-3248), harvested credentials, pivoted to a production MySQL/Nacos server, self-corrected a failed step in 31 seconds, then encrypted 1,342 config items with an ephemeral (never-saved) AES key, making the ransom demand unpayable-but-unrecoverable. Payloads were "self-narrating" with natural-language reasoning comments — strong evidence of LLM authorship. Cited here as the reference case for why agent-security tooling (guardrails, egress control, credential scoping) above matters in production.
• Lineation.ai - 🆕 ⚠️ July 2026 (new vendor). Agent accountability layer — observability, governance, and defense with forensic reasoning lineage. Aims to prevent goal hijacking, memory poisoning, and tool misuse; audit trail for SOC 2 / HIPAA / EU AI Act compliance. Cloud (free-to-start) and on-prem.
• First Recon AI Security Runtime - 🆕 July 2026. Enterprise AI governance platform that inspects every AI interaction (human-to-model, agent-to-tool, agent-to-agent) with a proprietary Semantic Security Engine — applies policies before data reaches a model and captures a full decision audit trail. macOS + Windows endpoint agent for governing AI use on device.
• CrowdStrike Falcon AIDR - GA December 2025. AI Detection and Response — visibility into employee AI use and agent activity across an enterprise, risk scoring, behavioral anomaly detection, prompt-injection blocking, and real-time policy enforcement at the AI interaction layer.
• Darkmoon - 🆕 Open-source (GPL-3.0) autonomous AI penetration-testing platform — an orchestrator LLM drives specialist sub-agents across web, API, Active Directory and Kubernetes targets, executing real exploits to produce proof-based findings rather than unvalidated alerts. Runs fully local and ships as both an MCP host and an MCP server. Notable design choice: a privacy gateway replaces real IPs, hostnames, credentials and paths with deterministic placeholders before anything reaches the LLM, rehydrating them locally, so target details never leave your perimeter.
• Exabeam Agent Behavior Analytics - 🆕 2026. Extension of Exabeam's behavior-intelligence platform to cover agentic AI risks — continuous verify-observe-analyze-improve loops instead of static guardrails.
• RufRoot / CVE-2026-59726 - 🆕 ⚠️ Disclosed to maintainers June 30, 2026; reported July 29, 2026. A CVSS 10.0 flaw in Ruflo (formerly Claude Flow), an open-source multi-agent orchestration layer for coding agents: the project's previous default Docker Compose config exposed Ruflo's MCP bridge to the network with no authentication, so a single request could invoke terminal_execute and reach all 233 tools behind the bridge — leaking LLM provider API keys and stored conversations. Worst part: attackers could write to AgentDB, Ruflo's persistent agent memory, so poisoned instructions survive the upgrade. Maintainers fixed the default in 24h (3.16.3), but recovery requires credential rotation and an AgentDB audit — patching alone is not enough. Found by Noma Labs. The canonical example of why agent memory is now part of your attack surface.
• Claude Code symlink exfiltration (Tego AI) - 🆕 ⚠️ July 24, 2026. A repo-committed CLAUDE.md with an @import pointing at a symlink can make Claude Code read files outside the project and fold their contents into its very first request — no tool call, no approval prompt, no warning, because the out-of-project read check validated the in-repo link path rather than what it resolved to. Reported via HackerOne; Anthropic closed it "Informative" on the grounds that the trust boundary is the initial folder-trust dialog. Worth reading before you let an agent loose on an untrusted repo.
• LLM Reasoning Model Jailbreaking (97% success rate) - 🆕 ⚠️ 2026 research. Published studies demonstrate large reasoning models can systematically jailbreak other AI systems with near-97% success rate and minimal human intervention. DarkLLM and commercial jailbreak frameworks have emerged on underground markets. Underscores why agent-to-agent trust boundaries need cryptographic enforcement, not just prompt-level guardrails.
• AI Red-Teaming Milestones (Claude + GPT-5.6 security evaluations) - 🆕 ⚠️ July 2026. Multiple frontier models in red-team evaluations demonstrated autonomous lateral movement: a pre-release GPT-class model executed 17,000+ automated actions in a single weekend, traversing Hugging Face internal clusters via a third-party package zero-day and exfiltrating evaluation answers. In separate tests, Claude models exploited weak credentials to breach three external organizations during cybersecurity challenges. Industry dubbed these the "software supply chain Chernobyl moment" for AI security evaluations.
• CrowdStrike 2026 Threat Hunting Report - 🆕 ⚡ 2026-08-03. AI-agent-triggered detections are 2.5× higher than human-initiated leads; Chinese APTs exploit PoC vulnerabilities within 24 hours of disclosure; STARDUST CHOLLIMA poisoned 300+ AI framework dependencies in a single day; a single LLMJacking campaign sent 200,000 API requests in 2 minutes.
• Straiker AI Runtime Security - 🆕 2026-08 (BH2026 showcase). AI-native agentic security platform — asset discovery (Discover AI), adversarial red-teaming (Ascend AI), runtime blocking (Defend AI). Blocks prompt injection, memory poisoning, identity abuse. Raised $85M total ($64M Series A, 2026-06).
• EU AI Act Article 50 Enforcement Live - 🆕 ⚡ 2026-08-02. ⚠️ Transparency obligations now enforced: AI conversational agents must identify as AI; generated content must carry machine-readable markers. Legacy system grace period until 2026-12-02. Max penalty €15M or 3% of global revenue.

🔍 RAG & Knowledge

Retrieval-augmented generation and knowledge management systems for agents.

• Oracle OCI Enterprise AI updates - 🆕 June 2026. Enterprise deployment of Cohere Rerank 4 to enhance RAG and agentic enterprise search, plus expanded support for new Alibaba/Google models.
• LlamaIndex - Data framework for LLM-based applications — ingest, structure, and access private data.
• Haystack - End-to-end LLM framework for building RAG pipelines and search systems.
• Unstructured - Open-source components for pre-processing documents for LLMs and RAG.
• Chroma - AI-native open-source embedding database.
• Weaviate - Open-source vector database for AI-native applications.
• Qdrant - High-performance vector similarity search engine and database.
• Pinecone - Managed vector database for high-performance AI applications.
• Milvus - Cloud-native vector database for scalable similarity search.
• RAGFlow - Open-source RAG engine based on deep document understanding.
• Docling - Document parsing and conversion for RAG and generative AI.
• Kotaemon - 🆕 Open-source RAG-based tool for chatting with documents.
• LightRAG - 🆕 Simple and fast RAG engine with graph-based knowledge indexing.
• R2R - 🆕 Production-ready RAG engine with built-in auth, observability, and ingestion.
• Vanna - 📦 Archived (2026-03). RAG for SQL — chat with your database using natural language.
• Morphik - 🆕 Multimodal RAG engine for documents containing tables, figures, and charts; rapidly-rising 2026 alternative to LlamaIndex for complex PDFs.
• Cognee - 🆕 Memory + reasoning engine that builds a knowledge graph as agents ingest documents; 2026 darling for "long-running research agent" stacks.
• RAG-Anything - 🆕 All-in-one multimodal RAG framework from HKU Data Science Lab. Built on top of LightRAG; concurrent pipelines for parallel text + multimodal processing; queries documents that interleave text, diagrams, tables, and formulae. MIT, 21K+ stars.
• A-MEM - 🆕 Agentic Memory system for LLM agents — dynamic organization of memories using Zettelkasten-inspired note linking; enables more flexible retrieval than static vector stores.
• LangChain Retrievers - LangChain's collection of retrievers and document loaders — the most widely used glue layer between raw sources and a RAG pipeline.
• Milvus 3.0 - 🆕 v3.0.0 tagged July 29, 2026 (public beta was May 2026). A "lake-native" architecture shift for the large-scale vector database — External Collections that query Parquet / Lance / Iceberg tables directly in S3/GCS/Azure object storage with zero-copy access, a manifest-based Storage V3 columnar engine, Spark DataSource V2 integration, runtime schema evolution, TEXT as a first-class type, and multi-vector StructList for late-interaction (ColBERT-style) retrieval.

💻 Coding Agents

AI-powered coding assistants and autonomous software engineering agents.

Terminal & CLI Agents

• Claude Code - ⚡ Anthropic's agentic coding tool. 80.9% SWE-bench score, handles complex multi-file bugs. Latest: v2.1.233 (August 14, 2026) — near-daily releases; recent additions include GitLab merge request support in --worktree, Bash tool memory cgroup limits, and a WebFetch cache TTL env var. May 2026 (v2.1.128–2.1.141) added the /goal command for cross-turn completion conditions, agent view, plugin loading from .zip archives + URLs, Ctrl+R global history search, broader MCP/hook handling, and enterprise feedback surveys.
• Codex CLI - OpenAI's open-source terminal coding agent (Rust, Apache-2.0, 82K+ stars). 77.3% Terminal-Bench score. May 2026 adds Codex Chrome extension for in-browser DevTools workflows, codex remote-control headless app-server, plugin-detail bundled-hook display, and Codex on Mobile preview (May 14) that lets ChatGPT iOS/Android remote-control the macOS Codex app. July 9–10, 2026: Codex integrated into the ChatGPT desktop app (macOS/Windows) — the standalone CLI continues with frequent point releases (v0.144.x, mid-July); Multi-Agent V2 runtime work landed in June CLI releases, and the desktop launch added a beta multi-agent feature with GPT-5.6. Codex Micro ($230 keypad for Codex control) announced July 15. ⚡ Latest stable CLI: v0.147.0 (August 7, 2026).
• Codex Security - 🆕 March 2026. Application-security agent that finds and fixes software vulnerabilities; available to OSS maintainers via the Codex-for-OSS program.
• Aider - 💤 AI pair programming in your terminal — works with any LLM, with first-class git commit handling. Last release Feb 2026 (v0.86.2).
• Goose - Open-source agentic coding CLI from Block — extensible, MCP-native, works with any LLM.
• Gemini CLI - 🆕 Google's terminal-first coding agent for large-context refactors.
• OpenCode - Open-source terminal AI coding agent (opencode.ai, 180K+ stars) — build/plan agents, LSP, MCP, desktop app in beta; unrelated to the archived opencode-ai/opencode.
• Crush - Terminal AI coding agent from Charm — successor to the archived opencode-ai/opencode; multi-model, LSP + MCP support.
• Grok Build - 🆕 May 25, 2026 (early beta). xAI's agentic CLI coding agent powered by grok-code-fast-1. Parallel sub-agents in isolated environments, daily release notes; available to SuperGrok and X Premium Plus subscribers. xAI's reply to Claude Code and Codex CLI. ⚠️ July 2026 reports found Grok Build uploading entire git repos to xAI storage — review before use on private code.
• Antigravity CLI - 🆕 May 19, 2026 (Google I/O 2026). Lightweight CLI companion to Antigravity 2.0 — create and interact with Google agent harnesses directly from the terminal. macOS / Linux / Windows. Reported to replace Gemini CLI for hosted-plan users from June 18, 2026 (the open-source gemini-cli repo remains active, 105K+ stars).
• Kimi Code CLI - 🆕 🇨🇳 June 6, 2026. Moonshot AI's terminal coding agent (TypeScript, MIT). Built-in coder / explore / plan sub-agents in isolated contexts, conversational MCP setup via /mcp-config, npm install. Aimed squarely at next-gen Kimi K2.6 agents.
• MAI-Code-1-Flash in GitHub Copilot - 🆕 Build 2026 (June 2, 2026). Microsoft's first fully in-house 5B coding model lands as a model picker option in GitHub Copilot — outperforms Claude Haiku 4.5 on four core coding benchmarks (SWE-Bench Pro 51.2% vs 35.2%) at significantly lower cost.
• Claude Agent SDK - SDK for building agents on the Claude Code harness (rebranded from Claude Code SDK in late September 2025). June 2026 Claude Code releases add hierarchical subagent spawning, Dynamic Workflows (fan out tens–hundreds of parallel subagents), rubric-driven Outcomes, and fallback model chains — powering the "ultracode" mode enabling up to 1,000 subagents per execution. Python + TypeScript.
• ai-delivery-spec - 🆕 ⚠️ Unverified. Spec-driven delivery framework for PMs working with AI coding agents (Claude Code, OpenClaw, Codex, Cursor, Copilot). 4 delivery tiers (Lite/Standard/L2/Full), 0D triage routing, prototype testability rules, AI runtime governance, 5 domain modules. SKILL.md convention; hosted on ClawHub.
• Ralph Harness - 🆕 ⚠️ Unverified. Tiny Python scaffold for guarded Claude Code/Codex/Gemini loops with repo-local specs, fresh-context iterations, git-hook gate, CI verification, and coverage gates. Installable via uvx ralph-harness demo. MIT.
• Amp - 🆕 ⚡ Sourcegraph's frontier coding agent (VS Code extension + CLI). No BYOK — model access is bundled, and its model-agnostic "Dial" router picks the model for you. July 2026 was a heavy month: paid subscriptions launched in beta July 18 (Megawatt $20/mo, Gigawatt $200/mo, with the option to attach your own ChatGPT or X Premium+/SuperGrok sub), self-scheduling agents July 21, "Multiplayer" shared-thread collaboration July 22, and event-driven Orbs July 23 — agents that wake on an external event (CI failure on GitHub, a new Linear issue, a monitor alert, a Discord message; anything that can send an HTTP request). August kept pace: "Attach Anything" uploads (video/logs/PDF/datasets, Aug 4), "Portals into Orbs" live-reloading previews (Aug 6), Dial running on a linked ChatGPT subscription (Aug 10), and Global Plugins and Skills (Aug 11). Closed-source.
• ZCode - 🆕 🇨🇳 July 2026 (ZCode 3.0). Z.ai's official agentic development environment for GLM-5.2 — a desktop app (macOS / Windows / Linux) wrapping file manager, terminal, Git panel, and live browser preview around an agent that plans, codes, reviews, and deploys. Also drives Anthropic and OpenAI models. Free tier with a daily token allowance; GLM-5.2 access via the paid GLM Coding Plan (Lite / Pro / Max).

IDE-Based Agents

• Cursor 3.11 - 🆕 ⚡ July 10, 2026. Side Chats (parallel AI conversations without disrupting main workflow), conversation history search, redesigned project/repo pickers, new Cloud Agent Hooks (granular control + observability over agent conversations — prompts, responses, sub-agent activities). August 2026: Google Workspace plugins let agents access Gmail/Drive/Calendar (Aug 3), and pre-built dev environments make Cloud Agents start 3× faster (Aug 13).
• Cursor 3.4 (Teams + PR review) - 🆕 May 11–13, 2026. Microsoft Teams integration (@Cursor in Teams delegates to cloud agents), faster parallel-agent plan execution, multi-repo / Dockerfile-based dev-environment configs for agents, /multitask async sub-agents, Vulnerability Scanner, granular per-model access controls.
• Cursor 3.3 - 🆕 May 2026. PR-review experience, parallel agents, enterprise model controls; previous 3.1 in April.
• Cursor SDK - 🆕 April 29, 2026 (public beta). TypeScript SDK exposing Cursor's runtime, harness, and models so developers can build programmatic agents on top of the Cursor stack — sandboxed cloud VMs, subagents, hooks, token-based pricing.
• Kilo Code - Open-source AI coding extension (VS Code / JetBrains) with Auto Model routing across 500+ models; acquired by Anaconda (2026). MiniMax models heavily featured.
• Cursor - The AI code editor with Feb 2026 update supporting up to 8 parallel agents.
• Windsurf → Devin Desktop - 🆕 Rebranded June 2, 2026. Cognition renamed the Windsurf IDE to Devin Desktop (windsurf.com now redirects to devin.ai): Devin Local (Rust rewrite, ~30% more token-efficient, subagent support) replaces Cascade, an Agent Command Center Kanban becomes the default surface, and it ships open Agent Client Protocol (ACP) support. Cascade reaches end-of-life July 1, 2026.
• Cline - Autonomous coding agent in your IDE — VS Code extension.
• Roo Code → Roomote - ⚠️ Discontinued as an IDE extension. Roo Code announced (April 22, 2026) shutdown of its VS Code extension, Cloud, and Router on May 15, 2026, pivoting to Roomote, a cloud coding agent (Slack/GitHub/Linear → PRs); roocode.com now redirects to roomote.dev.
• Void - 📦 Archived (repository archived on GitHub as of 2026-08; maintainers exploring new coding ideas; no further updates expected). Fork of VS Code positioned as the open-source Cursor alternative; data stays with you, BYO model.
• Continue - Open-source AI code assistant for VS Code and JetBrains.
• GitHub Copilot - Agent mode with expanded model access and gh copilot shell integration in early 2026. July 2026 updates: /security-review command (public preview) for AI-driven vulnerability scanning of in-progress changes; expanded "long-distance next edit suggestions" across full active file; C++ modernization agent GA in Visual Studio; refreshed Copilot Usage window for real-time billing tracking. July 31, 2026: Gemini 2.5 Pro and Gemini 3 Flash deprecated in all Copilot experiences (migrate to Gemini 3 Flash Exp or other available models); new Visual Studio .NET/Azure-specific agent (Copilot SDK); enterprise admins gain user-level model policy targeting (public preview).
• Kiro - AWS autonomous agent. Spec-driven development, manages up to 10 simultaneous tasks.
• Amazon Q Developer - AI coding companion deeply integrated with AWS ecosystem.
• Visual Studio 2026 Agent Mode + Skills - 🆕 VS 2026 Insiders May 12-15, 2026. Copilot Chat "Agent Mode" now ships a guided Skills workflow inside Visual Studio 2026: discover, manage, and author reusable Copilot Skills with whole-solution context, plus terminal command execution and tool invocation.
• JetBrains Rider AI Test-Writing Skill - 🆕 May 22, 2026. New AI Assistant skill for JetBrains Rider that surfaces .NET coverage data to Claude Code / Codex so agents target untested branches, reducing AI cost for test generation.
• Agent Skills (addyosmani) - 🆕 ⚡ August 2026 (v0.6.7, Aug 14). Addy Osmani's production-grade engineering skills for AI coding agents — skills encode the workflows, quality gates, and best practices that senior engineers use (DEFINE → PLAN → BUILD → VALIDATE), packaged so agents follow them consistently across every development phase. MIT license; 86,000+ stars.
• Cursor Router - 🆕 July 2026. Cursor's intelligent model-routing system analyzes each request and routes to the optimal model (Intelligence / Balance / Cost modes); integrates Grok models trained on trillions of Cursor interaction tokens (Grok 4.6 became the Cursor default on August 12, 2026). Companion: Cursor iOS app (July 2026) for mobile development. Part of Cursor IDE.
• Devin Desktop (formerly Windsurf) — July 2026 updates - 🆕 July 2026 updates. GPT-5.6 / Claude Opus 5 / Claude Fable 5 model support, Devin Outposts (run Devin agent on any machine), Agentic MapReduce architecture for distributed reasoning across large codebases. Acquired Poke (AI texting assistant) July 23, 2026.
• JetBrains Rider 2026.2 - 🆕 July 22, 2026. Enhanced AI agent intelligence and native GitHub Copilot integration; improved AI-assisted debugging and refactoring within the IDE ecosystem.
• Android Studio Quail 2 - 🆕 July 2026. Redesigned Agent Mode for AI-assisted Android development — memory leak detection, AI-powered crash analysis, and intelligent app-building workflows.

Autonomous Software Engineers

• Cursor 3.4 Cloud Agent Environments - 🆕 May 13, 2026. New dev environments for cloud agents: multi-repo workspaces, Dockerfile-based config with build secrets, 70% faster cached image layers, per-environment version history with rollback, audit logs, scoped egress and secrets. Companion to the Cursor 3.4 release.
• Devin Stacked PRs - 🆕 2026-07-30. Devin + GitHub: large tasks auto-split into independent small PRs, downstream PRs auto-rebased, focused context in Devin Review. Includes Faros AI data from 10,000+ developers.
• Devin Security Swarm - 🆕 July 1, 2026. Cognition's parallel-agent security product: finds vulnerabilities across a codebase, validates exploitability at runtime, and opens remediation PRs; found 36/50 real-world vulns at ~30% lower cost per finding than the next-best tool.
• Devin 2.2 - 🆕 February 24, 2026. End-to-end testing with computer use (Linux desktop + screen recordings), self-review/auto-fix before PR, 3× faster startup. Cognition's flagship autonomous software engineer (Devin 2.x line; Core plan from $20/mo since Devin 2.0). August 2026: Cognition reported to be raising at a $40B+ valuation as Devin approaches $1B in annualized revenue (press reports).
• OpenHands - Open-source platform for AI software developers as autonomous agents.
• SWE-agent - Turn LLMs into software engineering agents that fix real GitHub issues.
• Devika - 💤 Stale (no commits since 2025-09). Agentic AI software engineer — open-source alternative to Devin.
• GPT Engineer - 📦 Archived (2026-04). Specify what you want built, AI asks for clarification, then builds it. Foundational project of the autonomous-coding era, kept for historical reference.
• Codegen - 📦 Archived (2026-07-16). Programmatic code manipulation and multi-file refactoring SDK. Kept for historical reference.
• Qodo - 🆕 AI Code Review Platform focused on quality, security, and test generation.
• Google Antigravity 2.0 - 🆕 May 19, 2026 (Google I/O 2026). Standalone desktop application (macOS / Linux / Windows) for orchestrating multiple agents in parallel. Adds scheduled cron-style runs, async long-running tasks, dynamic sub-agents, and integrations with AI Studio / Android / Firebase. Companion Antigravity SDK lets you host the harness on your own infra; enterprise edition lands inside Gemini Enterprise Agent Platform.
• ChatGPT Work - 🆕 July 9, 2026. OpenAI's multi-step autonomous work agent — integrates with files and apps to complete entire jobs: spreadsheets, slide decks, documents, and small web apps. Desktop-app-centric with Chat/Work mode split. Accompanied by GPT Voice (July 23, 2026) for voice-directed task delegation.
• Cursor iOS - 🆕 July 2026. Cursor's mobile app for iOS — enables development work on the go with full model access and project sync.
• Cursor iPad + Agent Hooks - 🆕 July 28–29, 2026. Native iPad app (paid plans) with sidebar multi-agent monitoring, split-view code review, Apple Pencil annotations, and touch-optimized interface. Cloud Agent Hooks (GA) let developers observe agent reasoning and build self-correcting loops; "Cursor Start" regional pricing tier for India market.
• Claude Cowork - 🆕 ⚡ August 3, 2026. Anthropic's non-developer-facing autonomous work agent — web + mobile, connects to Slack and other business apps; delegates long-running tasks asynchronously. Companion to Claude Tag (Slack-native integration). Team and Enterprise plans.
• Claude Tag - 🆕 ⚡ August 3, 2026. Replaces the legacy Claude in Slack integration. Channel-level shared agent identity (@Claude), cross-session persistent context, asynchronous multi-day work. Forces migration from old Slack app; Team/Enterprise plans required.
• Prime Agent - 🆕 ⚡ August 2026 (v0.7.2). PrimeIntellect's open-source, self-improving coding agent that runs continuous RL feedback loops to improve its own capabilities over time. Designed for long-running autonomous coding workflows; MIT license; 14,500+ stars.

🤖 Physical AI & Embodied Agents

AI systems that perceive, reason about, and act in the physical world — humanoid robots, factory automation, Physical AI infrastructure. The next wave after language agents. 2026 H1 humanoid robot startups raised a record $8.6B globally.

Foundational Models & Research

• Microsoft physical-ai-toolchain - 🆕 June 2026. Production-ready open-source toolchain integrating Microsoft Azure cloud services with NVIDIA's physical AI stack. Agentic workflows for data curation, training, and deployment of autonomous mobile robots and manipulators.
• PhyAgentOS - 🆕 Self-evolving embodied AI operating system built on agentic workflows. Decouples cognitive functions from hardware for cross-platform robotic deployment with built-in auditability.
• ENPIRE - 🆕 June 2026. NVIDIA/CMU/UC Berkeley framework enabling AI agents to conduct their own robotics research — managing dual-arm robots, modifying algorithms, and training policies without human intervention.
• Kairos - 🆕 June 15, 2026. ACE ROBOTICS open-sources its Kairos world model — first world-model approach to top VLA systems on RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot, and DreamGen; weights on GitHub / Hugging Face / ModelScope.
• DYNA-2 - 🆕 ⚡ August 2026. Dyna Robotics' world-action model pre-trained on over 1 million hours of egocentric human video (~170 years) with a human-to-robot transfer scaling law — zero-shot pass rate at customer sites of 87% vs 46% for Dyna-1. Enables robots to complete end-to-end workflows in hospitality, logistics, and factory operations shift after shift; every deployment strengthens the fleet and scales from one site to hundreds.

• NVIDIA Cosmos 3 - 🆕 June 2026. Open world foundation model that unifies vision reasoning, multimodal generation, and action prediction so robots, AVs, and vision agents "think before acting" — rather than a text-only model bolted onto a controller. NVIDIA reports it as the top-ranked open VLM on VANTAGE-Bench (smart-infrastructure scene understanding) and the TAR traffic-anomaly challenge, with Cosmos 3 variants leading Artificial Analysis open-weights boards and topping Physics-IQ / R-Bench / PAI-Bench for world generation. Released under the Linux Foundation's OpenMDW 1.1 license — a single model-centric license covering weights, architecture, docs, datasets, benchmarks, and code. Available on build.nvidia.com, Hugging Face, and as NIM microservices. Includes Cosmos 3 Edge (4B params) for on-device deployment; launched alongside a Cosmos Coalition with FANUC, Fujitsu, and Sony Group.

• Google Gemini Robotics-ER 1.6 - 🆕 April 14, 2026. Robotics AI model with enhanced spatial/embodied reasoning (incl. analog instrument reading), available via the Gemini API.

• Google Gemini Robotics 2 - 🆕 July 30, 2026. Next-generation robotics AI suite from Google DeepMind — whole-body coordinated control, dexterous fine manipulation (knots, zippers), new leg locomotion capabilities; includes Gemini Robotics ER 2 as the "high-level brain" model that adapts to new robot embodiments within a few hours of adaptation, typically with fewer than 200 examples. Closed research; available via select partnerships.
• Project Prometheus (Bezos) - 🆕 💰 June 11, 2026. Jeff Bezos co-led Physical AI venture raised $12B at a $41B valuation to build an "artificial general engineer" for the physical world.
• NVIDIA Isaac GR00T - NVIDIA's foundation model platform for humanoid robots. Unveiled at GTC, expanded at Hannover Messe 2026.
• NVIDIA Industrial AI Cloud - 🆕 April 2026 (Hannover Messe). Deutsche Telekom-built AI factory infrastructure for industrial AI workloads.

Humanoid Robots

• Tesla Optimus V3 (Production Start) - 🆕 Production begins late July / August 2026 at Tesla Fremont, CA — first Optimus robots coming off an actual mass-production line. A second, larger assembly line is under construction in Texas targeting up to 10M units/year by 2027. V3 currently operates exclusively in Tesla factories (battery module handling, parts kitting); not yet available for public purchase. Evolved from Gen 2/Gen 3 platform — 37 joints, 1.2 m/s walking, 22-DoF hands.
• Figure 03 (Helix AI) - 🆕 Late 2025 announcement, ramping in 2026. First Figure model designed for the home: soft textile coverings, wireless charging, tactile sensors. May 2026 demo: two F.03 robots autonomously cleaning a room and making a bed in <2 minutes via visual coordination only.
• Figure 04 - 🆕 May 13, 2026. Founder Brett Adcock announces Figure 04 design finalized; component deliveries underway. Successor to F.03 with the Helix VLA model.
• Helix 02 package-sort 72h run - 🆕 May 13-16, 2026. Live-streamed Figure F.03 fleet runs Helix 02 fully autonomously on a package-sort line — ~22K packages on day one, then ~30K in the first 24 hours, ending in a stress test that hit ~88K packages over ~72 hours before mechanical failure. First public continuous-run evidence for a home-form-factor humanoid stack.
• Figure F.03 vs human 8-hour sort challenge - 🆕 May 18, 2026. Figure runs the first public head-to-head: one F.03 robot vs one trained human, 8-hour shift on the same package-sort line. Human wins narrowly — 12,924 parcels (2.79 s/item) vs 12,732 parcels (2.83 s/item). Tightest published gap to human throughput on a real industrial task to date.
• Boston Dynamics Atlas 100-lb manipulation + Hyundai 25K plan - 🆕 May 18-19, 2026. Boston Dynamics publishes video + technical blog showing Atlas lifting and carrying >100 lb loads (mini-fridge / washing machine) via RL + large-scale sim training; whole-body control adapts to weight shifts without per-object identification. Hyundai Motor Group commits to deploying 25,000+ Atlas units across Hyundai/Kia plants starting 2028 in Georgia.
• Hyundai completes full Boston Dynamics buyout - 🆕 💰 Late June 2026. Hyundai Motor Group buys SoftBank's remaining 9.65% stake for $325M, taking 100% ownership at a ~$3.4B valuation. New 5th-gen Atlas made a public appearance at the FIFA World Cup (July 5, 2026); commercial availability to external customers planned 2027, Georgia Metaplant deployment 2028.
• Boston Dynamics Atlas Gen 5 - 🆕 2026. Fifth-generation Atlas — "almost an order of magnitude" less complex than prior generations; lower production cost makes it more deployment-accessible. First deployments: Hyundai Robotics Metaplant Application Center and Google DeepMind.
• Figure 03 × BMW - 🆕 June–July 2026. Figure AI's Figure 03 humanoid robot deployed in BMW's Spartanburg, SC manufacturing plant for logistics tasks — first major commercial deployment of Figure hardware in automotive production.
• Unitree G1 deployed at JAL Haneda - 🆕 May 2026. Japan Airlines starts a Haneda ground-operations trial with Unitree G1 humanoids (baggage loading, container transport, cabin cleaning) — marketed as the first commercial airline trial of bipedal robots in active aviation service. US Congress separately moves to add Unitree to the entity list on national-security grounds the same week, underscoring how fast the embodied-AI supply chain is becoming geopolitical.
• Figure 02 + Helix 02 - 🆕 January 2026. Helix 02 expands whole-body autonomy (load/unload dishwashers, fold laundry); BotQ facility rated for 12K units/year.
• Unitree G1 + H2 - 🆕 CES 2026. G1 dance/boxing/skating demos, autonomous kung fu (February), and the 31+ DoF H2 humanoid.
• Unitree R1 Air - 🆕 Consumer humanoid at $4,900 — runs, flips, walks on hands.
• Unitree Gen 2 (lifelike skin) - 🆕 Realistic human-like skin with embedded pressure / temperature / touch sensors.
• Unitree GD01 - 🆕 May 2026. Nearly 10-foot manned mecha; pilot-driven, switches between bipedal and quadrupedal modes. Priced from ¥3.9M (~$650K). Tracks how the embodied-agent stack is starting to fork into operator-piloted form factors.
• Honor (荣耀) Humanoid - 🆕 April 19, 2026. Won the Beijing E-Town humanoid half-marathon in 50:26 — beating the human world record for the distance.
• Zhiyuan (智元) AGIBOT - 🆕 🇨🇳 April 17, 2026 (APC 2026). Declares 2026 "Deployment Year One"; 10,000th robot deployed, seven industry solutions, open-source AIMA architecture (Link-U OS, Genie Studio).
• Unitree H-series - Boston Dynamics competitor from China. Ongoing 2026 iterations.
• Unitree Shanghai IPO - 🆕 ⚡ 🇨🇳 August 2026. Unitree becomes China's first listed humanoid-robot maker — Shanghai STAR Market IPO priced at a reported ~$9B valuation, with the retail subscription oversubscribed thousands of times (Bloomberg: 5,526×) and gray-market pricing implying a sharply higher debut. A milestone for the commercial humanoid sector.
• 1X NEO (consumer humanoid) - 🆕 Pre-orders opened Oct 28, 2025, first US home deliveries in 2026. 5'6" / 66-lb home humanoid with 22-DoF hands, soft body, 4-hour runtime, on-device LLM, ~22dB noise. Early-access price $20,000 with $200 deposit, or $499/month subscription. Privacy "no-go" zones + face-blurring built in. First credible consumer-targeted humanoid to actually ship to homes.
• Mitsubishi Motors × Highlanders humanoid MOU - 🆕 July 9, 2026. Mitsubishi Motors signs an MOU with University of Tokyo startup Highlanders to co-develop and mass-produce humanoid robots (welding, logistics, engine assembly) at its Kyoto plant — reported target ~1,000 units/month, production from as early as 2027, addressing Japan's labor shortage.
• Agile Robots - German-Chinese robotics company building AI-driven industrial manipulation systems.
• Shenzhen Humanoid Pilot Line - 🆕 🇨🇳 Shenzhen launched its first pilot production line for humanoid robots on April 12, 2026 (Leju Robotics + Dongfang Precision in Longhua District). 2-hour assembly cycle, 500–1,000 units/year, with mass production moving to a 10,000-units/year Foshan facility.

Consumer Robotics & Wearables

• Doubao AI Glasses (ByteDance) - ⚠️ 🇨🇳 Gen 1 reportedly scrapped before launch ("too generic"); ByteDance fast-tracking a dual-model Gen 2 (July 2026) as the AI-glasses market grows ~130% YoY.
• Nothing AI Glasses/Earbuds - 🧪 Reported March 2026: Nothing plans AI smart glasses + earbuds, targeting a 2027 launch.
• Samsung Galaxy S26 (Gauss 2.3) - On-device agentic AI. Gauss 2.3 Think and Gauss O Flash variants.
• Meta Ray-Ban Display / Ray-Ban Meta - Meta's shipping AI-glasses line (incl. the Display model with Neural Band); Ray-Ban Meta Gen 3 rumored for late 2026.

Autonomous Driving

• Tesla FSD v14 - ⚡ Software update 2026.21.6 (August 10, 2026) bundles FSD v14.3.7 for HW4 cars alongside FSD v14.1 "Lite" for legacy HW3 (re-released Aug 11 after ECU-overheating issues); HW3 Lite expanding beyond US/Canada to South Korea. Unsupervised operation still limited to robotaxi programs.
• Waymo - ⚡ Continuing commercial L4 rollout in US cities through 2026. August 14, 2026: the CPUC cleared Waymo to operate paid driverless robotaxi service across 18 California counties (Bay Area, Sacramento, LA, Orange, San Diego and more) — its largest authorization to date, including freeways, rural roads, and night driving; rollout "gradual and guided by our safety framework."
• Pony.ai × Uber Europe - 🆕 ⚡ 🇨🇳 August 14, 2026. Pony.ai and Uber expand their partnership to deploy 2,000+ L4 robotaxis across five European cities plus the Middle East, building on the live Zagreb service — the jump from pilots to fleet-scale deployment.
• WeRide / Pony.ai / Baidu Apollo - 🇨🇳 Chinese L4 fleets expanding operational zones.

🎮 Agent Simulation & World Models

Research environments where agents are trained, observed, or stress-tested in simulated worlds. Increasingly relevant as world-model and embodied research bleeds into language-agent design.

• Generative Agents - 💤 Stanford's seminal Smallville simulacrum (Park et al., 2023). Memory + reflection + planning in a town of 25 LLM-driven characters. Reference implementation that influenced almost every multi-agent paper since.
• Voyager - 💤 First lifelong-learning agent in Minecraft — GPT-4 with skill library and curriculum (Wang et al., 2023). Still the canonical open-ended agent benchmark.
• SWE-Gym - Open environment to train SWE agents on real GitHub issues; companion to SWE-bench.
• WebArena - Realistic, reproducible web environment (Reddit / shopping / GitLab clones) used by OSWorld and most browser-agent papers. WebArena-Verified (ServiceNow, Dec 2025): all 812 tasks, reference answers, and evaluators manually reviewed; LLM-as-judge replaced with deterministic type-aware checks; 258-task "Hard" subset for cheaper evals.
• WorkArena - ServiceNow's enterprise workplace benchmark for browser agents.
• Genie 3 - Google DeepMind's interactive video world model — playable 3D worlds from a prompt. Closed-weights research, no public code.
• NVIDIA Cosmos - 📦 Archived. NVIDIA's foundation world model for embodied AI / robotics — generate physically plausible video futures. predict1 deprecated in favor of Cosmos-Predict2 (Predict 2.5 announced CES 2026); see also Cosmos 3 (June 2026) above.

• Snowflake Agent World Model (AWM) - 🆕 Open-sourced Feb 10, 2026; accepted to ICML 2026 May 1, 2026. Synthetic environment generation pipeline that ships 1,000 executable SQL-backed tool-use environments (35K+ tools, 10K tasks) exposed via a unified MCP interface — enables large-scale multi-turn agentic RL. Infrastructure merged into meta-pytorch/OpenEnv.

• Qwen-AgentWorld - 🆕 June 2026. Native language world model designed to simulate agentic environments (web, OS, terminal) by predicting the next state of the environment rather than the agent's next action.

• SimWorld - 🆕 Open-ended, realistic simulator built on Unreal Engine 5 for testing autonomous AI agents in complex physical and social environments.

📊 Benchmarks & Leaderboards

Standard evaluation suites and live leaderboards tracking frontier AI capability as of 2026.

⚠️ How to read the scores below. Leaderboard positions at the frontier now churn faster than any curated list can track, and mirror sites disagree with each other. While refreshing this section on 2026-07-30 we found three different "current" SWE-bench Verified leaders quoted by three reputable trackers on the same day, with an ~8-point spread. Rather than pick one and imply false precision, the numbers here are date-stamped snapshots of when they were read — always click through to the leaderboard before quoting a figure in your own work.

Two structural cautions that matter more than any single number: - SWE-bench Pro is compromised. OpenAI's 2026-07-08 audit of the 731 public tasks found ~27% flagged defective by an AI reviewer and ~34% by human engineers (over-strict tests, underspecified prompts, low-coverage tests). OpenAI withdrew its recommendation to use it as a primary coding metric. Top scores jumping 23% → 80% in eight months reflects benchmark noise as much as model progress. - Saturation makes small gaps meaningless. GPQA Diamond and SWE-bench Verified now sit in the mid-90s for several frontier models. A 0.5-point difference is within the noise of prompt scaffolding and retry policy; it is not a capability ranking. Build a small eval on your own repo or corpus — it will predict your results better than any public leaderboard.

• τ²-bench (tau2-bench) v1.0.1 - 🆕 2026-07 (v1.0.1). Sierra Research benchmark for AI agents in dual-control real-world scenarios (agent + simulated user operate the same system simultaneously). July 2026: banking_knowledge task errors corrected; older results not directly comparable.
• BenchLM - 🆕 Composite leaderboard that aggregates multiple benchmark families. July 2026 top: Claude Fable 5 91, Claude Mythos 5 89, Gemini 3.1 Pro 88 (79 ranked / 281 tracked models, 296 benchmarks).
• SWE-bench Verified - Real-world GitHub issue resolution benchmark. April 2026 top: Claude Mythos 93.9%, Claude Opus 4.7 87.6%. July 2026 top: Claude Mythos 5 leads at 95.5%, Claude Fable 5 at 95.0%.
• GPQA Diamond - 💤 Stale dataset repo (last update 2024-09). Expert-level science reasoning. April 2026 top: Gemini 3.1 Pro 94.3% (world-record), Claude Opus 4.7 94.2%.
• ARC-AGI 2 - Abstract reasoning over novel tasks. GPT-5.5 leads at 85% (July 2026); Gemini 3.1 Pro 77.1% (Mar 2026).
• ARC-AGI-3 - 🆕 🧪 Third-generation ARC benchmark — agents must adapt on the fly to novel interactive environments (Kaggle competition). Early top scores near zero (Gemini 3.1 Pro ~0.37%).
• OSWorld - Desktop GUI manipulation. Claude Fable 5 / Mythos 5 lead at 85% (July 2026); GPT-5.4 75% (exceeded human baseline).
• Arena (formerly LMArena / Chatbot Arena) - Crowdsourced chat preference battles. July 2026: Claude Fable 5 leads (Elo 1525), ahead of Opus 4.8 (1512) and GPT-5.5 Pro (1510).
• MMLU-Pro - Multi-task language understanding, harder successor to MMLU.
• LiveCodeBench - Contest-style coding benchmark, updated continuously to resist contamination.
• AIME 2025 / Humanity's Last Exam (HLE) - Elite math / PhD-level general reasoning.
• Terminal-Bench - CLI agent evaluation. Codex CLI 77.3%.
• Wolfram LLM Benchmarking Project - Code generation benchmark from English spec to Wolfram Language. Updated continuously.
• Terminal-Bench 2.0 - Late 2025 / early 2026. 89 curated terminal tasks (compile, train, configure, debug). Leaderboard top: NexAU-AHE + GPT-5.5 at 84.7% (May 14, 2026).
• GDPval - OpenAI economic-value benchmark (released Sept 2025, arXiv:2510.04374) across 44 occupations / 9 industries; 1,320 expert-built tasks. GDPval-AA leaderboard leader (July 2026): Claude Opus 4.8 (Elo 1890), ahead of GPT-5.5 (1769). Expanded/highlighted again at the GPT-5.6 launch (July 9, 2026).
• SWE-bench Pro - ⚠️ Do not use as a primary metric — see the audit note at the top of this section. Repository-level engineering successor to Verified. Claude Opus 4.7 64.3% > GPT-5.5 58.6% (Claude leads on long-horizon repo work). July 2026: Claude Mythos 5 tops at 80.3%; ⚠️ OpenAI audit (July 8, 2026) revealed ~30% of Pro tasks are broken (overly strict tests, underspecified prompts) — OpenAI has retracted its recommendation to use SWE-bench Pro as a primary coding-agent metric.
• LLM-Stats Live Leaderboard - 🆕 Continuously-refreshed cross-benchmark dashboard for newly-released models.
• τ²-Bench (Tau-Bench) - 🆕 Sierra Research's benchmark for tool-agent-user interaction in real-world domains (retail / airline). Measures multi-turn tool use, DB ops, and policy adherence. April 2026 leader: Claude Mythos Preview at 89.2% across 38 evaluated models. The repo now also hosts τ³-Bench 1.0.0 (Mar 18, 2026) adding voice + telecom/banking domains. MIT.
• Gartner Magic Quadrant 2026 — Enterprise AI Coding Agents - 🆕 2026. Gartner's first MQ for Enterprise AI Coding Agents. Leaders: GitHub Copilot, OpenAI Codex, Cursor (12 vendors evaluated; Tabnine a Visionary). Signals the coding-agent market reaching enterprise maturity.
• Terminal-Bench 2.1 - 🆕 June 2026. Updated Terminal-Bench with expanded CLI task coverage; leaderboard: Claude Code + Fable 5 leads at 83.8% (June 7, 2026), Codex + GPT-5.5 at 83.1%. Gap between top CLI agents now under 1%.
• Agent Memory Benchmark (AMB) - 🆕 Open benchmark for evaluating agent memory systems, released alongside Hindsight by vectorize.io. Tests retain / recall / reflect operations against long-term task performance.
• Agents' Last Exam (ALE) - 🆕 June 2026. UC Berkeley (RDI) + Snorkel AI benchmark evaluating agents on long-horizon, economically valuable professional tasks across 55 subfields — created with 300+ industry experts (public release ALE-V1: 147 reference tasks from a 1,500+ task corpus; arXiv:2606.05405). Launch pass rate ~2.6%; July 2026 leaderboard top: Codex + GPT-5.6 at 53.6/100 (highlighted at the GPT-5.6 launch). Living benchmark with task rotation every 6 months.
• JetBrains Kotlin Benchmark - 🆕 July 2026. Public benchmark from JetBrains measuring AI coding agents on real-world Kotlin development tasks — code generation, refactoring, test writing, and bug fixing. Gives developers a credible, language-specific evaluation beyond general SWE-bench scores.
• Stripe Agent Benchmark - 🆕 March 2026. End-to-end evaluation suite testing whether AI agents can build complete Stripe integrations from scratch — covering code generation, test execution, and production-style validation. Claude Opus 4.5 achieved 92% average score on full-stack API integration tasks in initial evaluations.
• GAIA Benchmark - General AI Assistants benchmark: multi-step reasoning + tool use + web browsing on 466 real-world tasks. July 15, 2026 public snapshot: Claude Mythos 5 and Claude Fable 5 lead at 52.3% (tied), followed by GPT-5.4 Pro at 50.5%. Designed to resist data contamination.

🖥️ Computer Use & Desktop Agents

AI agents that can see, control, and automate desktop environments at the OS level. For purely browser-based agents see 🌐 Browser & Web Agents.

• Perplexity Personal Computer for Windows - 🆕 July 28, 2026. Perplexity's multi-model agent orchestrator expands to Windows 10/11 — local files, native apps + Microsoft 365 suite, and cross-device workflows in one system. Pro/Max/Enterprise subscribers. Builds on the Mac version (April 16) and the hybrid local/cloud inference orchestrator previewed at Computex 2026.
• Claude Computer Use - Anthropic's computer-use capability — Claude sees the screen and uses mouse/keyboard to automate any software.
• ChatGPT Agent - Successor to Operator (deprecated 2025) — agent mode in ChatGPT for browsing, booking, form-filling, and web task automation.
• Google Project Mariner - 📦 Discontinued (May 2026). Browser-agent research project; capabilities merged into Gemini and Chrome.
• Microsoft Copilot Agents - 🆕 Autonomous background agents across the Microsoft 365 stack. Beyond sidebar — executes tasks and surfaces for approvals.
• Open Interpreter - A natural language interface for computers — let LLMs run code locally.
• Manus AI - 🆕 🇨🇳 Autonomous general-purpose AI agent with cloud-to-local hybrid model. Handles research, coding, and complex multi-step tasks.
• Genspark - 🆕 All-in-one autonomous work agent with mixture-of-agents architecture. Can make phone calls.
• Beam AI - 🆕 Self-learning desktop agents that refine logic based on successful outcomes.
• AICraft - 🆕 ⚠️ 🇨🇳 Unverified (created June 2026, sole maintainer, low traction — listed for completeness, vet before use). Windows desktop "AI capability launcher" that treats LLM skills, MCP tools, RAG sources and memory as hot-swappable modules — the pitch is loading them like Minecraft mods. One-click DeepSeek setup, per-role emotion portraits, three-tier memory, token billing; portable exe, no install. Python FastAPI + React 19 + ChromaDB, Apache-2.0.
• Microsoft Copilot Studio Computer-Using Agents - 🆕 GA May 13, 2026. Build agents inside Copilot Studio that interact directly with websites and desktop applications through the UI — Microsoft's first-party answer to Claude Computer Use, now generally available across Microsoft 365 / Power Platform deployments.
• ChatGPT Workspace Agents - 🆕 Research preview April 22, 2026; credit-based pricing May 6, 2026; EKM support May 7, 2026. OpenAI's successor to Custom GPTs for enterprises — cloud-side agents with file access, code execution, scheduled runs and built-in connectors for Slack, Google Drive, Salesforce. Available on Business / Enterprise / Edu / Teachers; powered by Codex.

🌐 Browser & Web Agents

Frameworks and infrastructure for agents that interact with the web through real browsers — navigate, click, scrape, and complete multi-page workflows.

• Cloudflare Kitesurf - 🆕 ⚡ August 6, 2026 (beta, Cloudflare Agents Week). Cloudflare's new serverless browser built specifically for AI agents — runs on Workers, stateless and isolated per session, optimized for token count and context-window efficiency over pixel-perfect rendering. Supports Puppeteer and Playwright; 3.1× less CPU and 4.7× less memory than Chromium for screenshot workloads, 215K+ Web Platform Tests passing. Free during beta via Browser Rendering. Trade-offs: no video playback, no WebGL, limited persistent-auth support.
• Browser Use - v0.13.7 (July 27, 2026). Make websites accessible for AI agents with browser automation. The de-facto open-source choice in 2026, 109K+ stars.
• Stagehand - The SDK for browser agents — typed act/extract/observe primitives over Playwright by Browserbase. MIT.
• Steel Browser - 🆕 Open-source browser API for AI agents — batteries-included sandboxed Chromium with session persistence and proxy rotation. Apache-2.0.
• Skyvern - Automate browser-based workflows with LLMs and computer vision. AGPL-3.0.
• AgentQL - Query language + Playwright integration for semantic web extraction. Reliable on dynamic, cluttered pages.
• Hyperbrowser MCP - 🆕 Hosted headless-browser fleet exposed as an MCP server — plug into Claude/GPT/LangChain via the standard tool interface.
• Playwright MCP - 🆕 Microsoft's official Playwright server exposed as an MCP tool. Production-grade automation primitives without rolling your own bridge.
• MultiOn - 📦 Hosted browser agent platform with native Reasoning + Memory; multion.ai now redirects to AGI, Inc. (theagi.company). Closed-source.
• Browserbase - Headless browser infrastructure built specifically for AI agents — stealth, persistence, captcha solving, observability.
• BrowserOS - 🆕 First open-source browser with built-in AI agents — privacy-first Chrome alternative. Natural-language task automation without coding; local-first design competes with Perplexity's Comet and Arc's AI features.
• Vercel Agent Browser - 🆕 Headless browser automation CLI for AI agents. June 2026 release adds vitals command for Core Web Vitals (LCP/CLS/TTFB/FCP), pushstate for SPA navigation, out-of-process plugin system, MCP server mode, and @agent-browser/sandbox for hosted environments. Apache-2.0, 37K+ stars.
• Google ADK — browser & A2A integration - Google's Agent Development Kit (v2.5.0, July 2026) — multi-agent composition, MCP tools, A2A Protocol integration.
• WebBrain - 🆕 July 2026. Open-source MIT browser extension (Chrome + Firefox) for automating web tasks with local or cloud LLMs. "Ask mode" for read-only summarization and data extraction; "Act mode" for clicks, form fills, and navigation. Local-first by design — data never leaves the device when using llama.cpp / Ollama.
• Muse Spark 1.1 (web agent) - 🆕 💰 July 9, 2026. Meta Superintelligence Labs' first paid agentic model, via the Meta Model API public preview — scores 69.0 on WebArena-Verified (behind Claude Opus 4.8's leading 71.2).
• Firecrawl v2 - 🆕 v2.11.0, June 2026. Major update to the agentic web scraping platform: improved JavaScript rendering, live crawl webhooks, and batch URL processing. 150K+ stars. AGPL-3.0 (MIT SDKs).
• Claude in Chrome - 🆕 Anthropic's browser-agent Chrome extension — Claude navigates, fills forms, and acts across tabs. ⚠️ July 2026 research showed rogue-extension prompt-injection risks; review permissions before use.
• Perplexity Comet - Perplexity's agentic AI browser with Comet Assistant (background agent); free tier since Oct 2025; Perplexity raised $200M for Comet in June 2026.
• Safari MCP Server - 🆕 July 1, 2026 (Safari Technology Preview 247). Apple's native browser-level MCP integration — Safari exposes browsing context, tab management, and page content to MCP clients. First major browser to ship native MCP support. Developer preview only.

🗣️ Voice & Multimodal Agents

Voice-enabled and multimodal AI agent platforms.

• ByteDance Seedance 2.5 - 🆕 ⚡ July 31, 2026. First large-scale AI video generation model supporting single-shot 30-second audio+video synthesis with multi-round extensions; up to 30 images + 10 video clips + 10 audio clips as multi-modal reference input in a single pass; local video editing. Rolling out on Jimeng AI and Doubao Pro; API via BytePlus ModelArk pre-release.
• xAI Grok Voice Think Fast 2.0 - 🆕 ⚡ 2026-07-29 (grok-voice-latest auto-upgraded from 2026-08-05). Next-gen speech-to-speech: first-byte audio latency 1.25s → 0.70s; transcription accuracy +1.4× across 24 languages; reasoning token usage −60%; $0.08/min.
• AgentLine - 🆕 ⚠️ Unverified. Telephony infrastructure for AI agents — provision phone numbers, make/receive calls, real-time transcription to JSON webhooks. Pitched as a thinner alternative to Twilio for agent voice pipelines; submitter claims 30+ paid users, no third-party adoption signal yet.
• ElevenLabs - AI voice platform with conversational AI agents and realistic speech synthesis. July 2026 update: Music Finetunes API (programmatic custom model management), per-agent sentiment analysis, nested agent transfers, RAG knowledge-base queries, auto-translated transcripts, faster generation with improved tonal consistency for long audio. $500M Series D at an $11B valuation (closed February 4, 2026, led by Sequoia; total funding >$781M), alongside crossing $500M ARR — the best-funded pure-play voice-AI vendor going into H2 2026.
• Vapi - Enterprise voice AI platform — build, test, and deploy voice agents. $50M Series B announced May 12, 2026 after crossing 1B platform calls; May 2026 updates ship Squads v2 (multi-assistant orchestration), Composer alpha (prompt-built agents), Simulations alpha (systematic AI-powered testing), and GA of the Soniox low-latency multilingual transcriber.
• Retell AI - Build production-ready conversational voice AI agents.
• Bland AI - AI phone calling platform — enterprise-grade conversational AI.
• Hermes - 🆕 ⚠️ Unverified (founders' beta). White-label voice agent platform aimed at agencies: agents, native CRM, outbound/inbound campaign orchestration, and per-client usage billing in one system. Sits a layer above raw voice APIs (Vapi/Retell) by bundling the CRM and billing an agency would otherwise assemble itself. Paid plans start at $149/mo with included minutes. No independent adoption data yet.
• LiveKit Agents - Build real-time multimodal AI agents with voice, video, and data.
• ByteDance SeedRealtime - 🆕 ⚡ 🇨🇳 August 5, 2026. Native audio-visual full-duplex LLM from ByteDance — continuously processes audio, video, and text streams, watching, listening, and speaking simultaneously in real time. Replaces traditional cascaded voice-agent pipelines; integrated into the Doubao app. No public API or model weights yet.
• Pipecat - 🆕 ⚡ v1.7.0 — August 1, 2026. Open-source framework for voice and multimodal conversational AI; v1.7.0 is the first stable 1.x release (deprecated 0.x APIs removed, Python 3.11+ required, universal LLMContext for mid-call model switching, filter_background_audio for ElevenLabs STT). BSD-2-Clause.
• Vocode - 💤 Stale (last release June 2024). Open-source library for building voice-based LLM agents.
• Bolna - End-to-end open-source voice AI agents framework.
• Cartesia - 🆕 Ultra-low-latency voice AI for real-time conversational agents.
• Meta Voice AI - 🆕 Former PlayHT/Play.ai team's tech, integrated into Meta AI, AI Characters, and Meta wearables after July 2025 acquisition. Original Play.ai platform shut down Dec 31, 2025.
• Sesame - 🆕 Voice AI companion with emotional understanding and natural conversation.
• ElevenAgents - 🆕 ElevenLabs' full-stack voice-agent platform (April-May 2026 updates): MCP, multimodal messages, conversation topic discovery, knowledge-base search, pre-tool speech controls. First voice-agent platform to earn AIUC-1 certification.
• Cartesia Line - Code-first voice-agent platform (launched Aug 2025) built on Cartesia's Sonic TTS + Ink STT with background reasoning and on-prem deployment options; ~40-90ms time-to-first-audio.
• Deepgram Voice Agent API - 🆕 Single endpoint bundling STT (Nova-3) + LLM routing + TTS (Aura-2) + Flux conversational STT with mid-call language switching across 10 languages.
• OpenAI Realtime API (GPT-Realtime-2) - 🆕 May 7, 2026. GPT-5-class reasoning over voice with parallel tool calls, 128K context; shipped alongside GPT-Realtime-Translate and GPT-Realtime-Whisper. Updated to gpt-realtime-2.1 / 2.1-mini on July 6, 2026 (better alphanumeric recognition, noise handling, lower latency).
• Dograh - 🆕 Open-source, self-hostable voice AI platform — an alternative to Vapi / Retell. On-prem, BYOK across Speech-to-Speech or LLM/STT/TTS, visual workflow builder, MCP-native, telephony support. BSD-2-Clause, 4K+ stars.
• Hume TADA - 🆕 March 2026. Hume AI's first open-source TTS — Text Audio Dual Alignment: text and audio generated in one synchronized 1:1 token stream; zero content hallucinations, ~0.09 RTF, TADA-1B/3B-ML (9+ languages), runs on a smartphone; powers the next generation of EVI voice agents. MIT code, Llama 3.2 licensed weights.
• OpenYabby - 🆕 Open-source macOS voice-driven multi-agent orchestrator — Realtime API + CLI runners + multi-channel orchestration. A lead agent plans the work and delegates to sub-agents for review and QA. MIT.
• Grok Voice Agent Builder - 🆕 July 1, 2026. xAI's no-code platform for production voice agents on Grok Voice — telephony with free provisioned numbers, knowledge collections, tools/MCP connectors, guardrails, 80+ voices plus ~2-minute voice cloning; $0.05/min in beta.
• GPT Voice - 🆕 July 23, 2026. OpenAI's voice interface for ChatGPT Work — powered by GPT-Live technology, allows users to direct multi-step agent workflows through natural voice commands.

📱 Personal AI Agents

AI agents designed for personal use, productivity, and daily life assistance.

• OpenClaw - 🆕 Personal AI agent platform with skills, memory, Dreaming, Canvas/A2UI, ACP coding harness integration. Runs on your machine with multi-channel messaging.
• Rabbit R1 - Dedicated AI hardware device with a large action model for personal assistance.
• Limitless - 📦 Acquired by Meta (late 2025); pendant sales discontinued. Personalized AI powered by what you've seen, said, and heard (formerly Rewind); team folded into Meta's AI-wearables effort.
• Open Interpreter - A natural language interface for computers — let LLMs run code locally.
• 01 Light - 💤 Stale (no commits since 2024-11). Open-source voice interface for computers.
• Leon - Open-source personal assistant — lives on your server.
• Khoj - Personal AI second brain — search and chat with your notes, docs, and images.
• Humane AI Pin - ⚠️ Discontinued Feb 28, 2025 (acquired by HP, device shut down). Originally a wearable AI device with a screenless, ambient computing experience.
• Arahi AI - 🆕 Personal productivity and business automation assistant.
• Lindy AI - 🆕 No-code AI agent for email, calendar, and workflow automation.
• MuleRun - 🆕 Always-on agents for recurring tasks and background automation.
• Gemini Intelligence - 🆕 May 12, 2026 (Android Show: I/O Edition). Proactive agentic AI features integrated into Googlebooks laptops, Wear OS, Android Auto, Android XR, and starting on the latest Samsung Galaxy + Pixel devices. Auto-creates shopping carts from grocery lists, books spin classes, filler-word removal via the Rambler speech-to-text.
• Gemini Spark - 🆕 I/O 2026 (May 19, 2026). Google's 24/7 autonomous agent in the Gemini app — runs multi-step processes proactively with Gmail/Workspace integration; expanded to a native Mac app July 1, 2026. July 2026: expanded from Gemini Ultra to Pro tier; handles scheduling, drafting, and inbox triage autonomously.
• Gemini Notebook - 🆕 July 2026 (rebranded from NotebookLM). Rebranded and upgraded: adds code execution, chart generation, and automatic source citation alongside the existing audio overviews and Q&A capabilities.
• QwenPaw - 🆕 🇨🇳 May 2026 rebrand from CoPaw. Self-hostable personal assistant in the Qwen / AgentScope family. Local-first memory, hot-loadable skills, multi-agent collaboration, multi-channel (DingTalk / Feishu / WeChat / Discord / Telegram), tool guard + skill scanner. Apache-2.0.
• AI Growth Agents for Marketers - 🆕 ⚠️ Unverified (early-stage). Growth marketing prompts and Python agents built from real fintech campaigns in Southeast Asia. Covers campaign briefs, MEU planning, and A/B test analysis with multi-agent workflows. Agent Skills format — installable via npx skills add. Bilingual VI + EN. MIT.
• Microsoft Scout - 🆕 Build 2026 (June 2, 2026). Microsoft's always-on personal agent built on the OpenClaw framework — proactive across cloud / desktop / web, connects to Teams / Outlook / OneDrive / SharePoint. Each agent runs under its own Entra identity with continuous policy-conformance checks and audit trails. Private preview via the Microsoft Frontier program; requires Intune policy + GitHub Copilot license.
• Lenovo Qira / Motorola Qira - 🆕 CES 2026 (Jan 6, 2026). Cross-device "Personal Ambient Intelligence System" co-developed by Lenovo and Motorola — context-aware AI that perceives, thinks, and acts across PCs / phones / tablets / wearables. Rolling out on select Lenovo devices in Q1 2026, expanding to Motorola phones thereafter; first major OEM ambient-AI play.
• Yao Agents - 🆕 🇨🇳 May 2026. Local-first AI execution platform with 30+ domain Experts (coding, writing, data analysis, PM) and autonomous Robot workers. Features a 5-stage Pipeline (Inspiration→Goals→Tasks→Validation→Delivery), Docker sandbox isolation, multi-platform messaging (WeChat/Feishu/DingTalk/Telegram/Discord), MCP support, BYOK model configuration, and Tai Link for cross-device agent orchestration. Open-source engine: YaoApp/yao.
• AgentArk - 🆕 🧪 June 2026 (v0.0.1, beta — not for production). Personal AI OS prioritizing local control and security; self-learning via the GEPA optimizer runtime.
• nanobot - 🆕 Ultra-lightweight open-source personal AI agent (41K+ stars). April 2026 releases (v0.1.5.x) added thread-scoped sessions, auto-compact memory, Dream consolidation, DeepSeek-V4 support, and Windows support.

📱 Mobile Agents

GUI agents that drive Android/iOS phones — the next frontier after desktop computer-use. Most major model providers now ship a mobile-grounded variant.

• Mobile-Agent - 🇨🇳 Alibaba's flagship multimodal phone-control agent family (v1 → v3, plus Mobile-Agent-E and Mobile-Agent-V). State-of-the-art on Android benchmarks.
• AppAgent - 💤 Tencent's multimodal agent that operates smartphone apps by tapping/swiping. Influential early implementation; succeeded by AppAgentX (Mar 2025).
• Apple Intelligence - On-device agent layer in iOS / iPadOS / macOS. App Intents and screen-aware actions across the OS.
• Samsung Galaxy AI / Bixby 2.0 - On-device Gauss-powered agentic capabilities baked into the Galaxy S26 line.
• Google Gemini for Android - Replaces Google Assistant on Android with full Gemini-powered, app-aware actions including system intents and Workspace.
• Magma - Microsoft Research foundation model for multimodal agents — grounds across UI, robotics, and physical action; targets phones, web, and embodied tasks.
• mobile-use - 🆕 Open-source framework (Apache-2.0, 2.5K+ stars) letting AI agents drive real Android and iOS apps as if they were a human — UI-aware navigation, natural-language control.
• agent-device (Callstack) - 🆕 February 2026. Lightweight, token-efficient CLI for automating iOS and Android devices + simulators. Command model designed for AI agents and CI; MIT, 2.6K+ stars.
• iOS 27 Siri AI (preview) - 🆕 Preview July 2026 (GA fall 2026). Apple's completely rebuilt Siri powered by Apple Intelligence — cross-app context awareness, natural language Shortcuts automation, multi-AI-model marketplace anticipated in iOS 27. Developer beta available July 2026.
• EU Android AI Openness Ruling - 🆕 July 17, 2026. European Commission orders Google to provide rival AI assistants deeper Android access — camera, microphone, and app-control APIs — paving the way for third-party mobile AI agents. Must be implemented by August 2027 in Android 18.

🏢 Enterprise Agent Platforms

Enterprise-grade platforms for deploying AI agents at scale.

• GPTBots.ai LoopAgent - 🆕 ⚡ 2026-08-03. Production-grade execution engine for enterprise AI agents: sandboxed code execution, lazy-loaded Skills, versioned System Identity Prompt Diff, seamless human-handoff context summaries. ⚠️ Unverified (GlobeNewswire announcement; primary URL unverified).
• Salesforce Agentforce 360 - Autonomous AI agents for enterprise CRM — sales, service, and marketing. Spring 2026 release ships Agentforce Builder (conversational agent authoring), Agent Script (deterministic behavior control), Agentforce Voice (Amazon Connect / Five9 / Genesys / NiCE / Vonage + SIP), and Intelligent Context on top of the new Data 360. Customers across 124 countries report ~85% autonomous query resolution.
• Microsoft Copilot Studio - Build and customize AI agents and copilots for your organization.
• Gemini Enterprise Agent Platform - 🆕 April 22, 2026 (Google Cloud Next '26). Evolution of Vertex AI into a unified hub for building, scaling, governing, and optimizing enterprise agents. Supports Gemini 3.1 Pro/Flash, Lyria 3, plus third-party models (Claude Opus/Sonnet/Haiku). Integrated agent DevOps, security, and orchestration.
• Google Vertex AI Agent Builder - Renamed April 2026 — Vertex AI's agent-building capabilities are now part of the Gemini Enterprise Agent Platform (see above): Agent Studio, Model Garden, Google Antigravity orchestration.
• Amazon Bedrock Agents - Build AI agents that can execute multi-step tasks across company systems.
• ServiceNow AI Agents - AI agents for enterprise IT service management with AI Control Tower. 🆕
• ServiceNow Action Fabric (MCP Server) - 🆕 May 5, 2026. ServiceNow opens its AI Platform to any AI agent (Claude, Copilot, custom) via a GA MCP Server bundled with every Now Assist and AI Native SKU. Every action runs through AI Control Tower — identity-verified, permission-scoped, audited; OAuth, consumption metering, role-based tool packages out of the box. Anthropic (Claude Cowork) is first design partner.
• IBM watsonx Orchestrate - AI assistant platform to automate work across enterprise applications.
• Oracle AI Agents - Enterprise AI agents integrated with Oracle Fusion Cloud ERP. 🆕
• Moveworks - Enterprise copilot platform — AI that works across every system. Acquired by ServiceNow (closed December 15, 2025).
• UiPath Agentic Automation - 🆕 Agentic reasoning layered onto RPA bot estates for intelligent process automation.
• AgentX - 🆕 Agentic enterprise solution for scalable AI automation with plug-and-play chatbots.
• Sistava - ⚠️ "AI employees on demand" for sales, marketing, support, recruiting, and operations — agents work inside your tools with persistent memory; from $19/month.
• Sema4.ai - 🆕 Enterprise AI agent platform with Python-first approach and built-in governance.
• SAP Business AI Platform + Autonomous Suite - 🆕 SAP Sapphire 2026 (May 12, 2026). SAP unveils the "Autonomous Enterprise": SAP Business AI Platform as the unified AI foundation; SAP Autonomous Suite adding agents to existing apps across finance, supply chain, procurement, HR, and CX; Joule Studio for building enterprise agents and agentic workflows; Joule Work UX; and seven Industry AI solutions. Claude is among the foundation models powering Joule agents.
• Microsoft Agent 365 + Microsoft 365 E7 - 🆕 May 1, 2026 GA with extended May rollouts. Identity-first control plane for governing and securing AI agents across enterprise environments; $15/user/month standalone, $99/user/month inside the new Microsoft 365 E7 "Frontier" suite. May 2026 update adds AWS Bedrock + Google Cloud registry sync, Intune/Defender preview policies, and SASE for agents.
• OpenAI Guaranteed Capacity (Compute Annual Pass) - 🆕 May 19, 2026. Long-term enterprise compute reservations (1 / 2 / 3-year terms, larger discounts at longer terms) sold as a structured product. Designed to derisk enterprise rollout of GPT-5.5-class agents — OpenAI's reply to the Anthropic Priority Tier model.
• Bristol Myers Squibb ↔ Claude Enterprise - 🆕 May 20, 2026. BMS standardises on Claude Enterprise as its shared intelligence platform for 30,000+ employees, embedding agentic Claude into drug-discovery / development / delivery pipelines. First top-5 pharma to make a public, company-wide Claude commitment.
• Kore.ai Artemis Agent Platform - 🆕 May 21, 2026 (launched on Azure). AI-native enterprise agent platform built around the new YAML-style Agent Blueprint Language (ABL) for declarative multi-agent workflows. Kore.ai's structural challenge to Copilot Studio and Agentforce.
• FPT Flezi Foundry™ - 🆕 May 21, 2026. AI-augmented delivery platform with two governed Service-as-a-Software modes — Agentic Development Lifecycle (ADLC) for full SDLC agent crews and Agentic Managed Services (AMS) for incident-resolution agents on top of existing ITOps.
• Amazon Bedrock AgentCore Payments - 🆕 May 7, 2026 (preview). Managed payments for AgentCore agents — autonomous payment for APIs, MCP servers, web content, and other agents via Coinbase (CDP wallet, x402 Bazaar) and Stripe (Privy wallet) integrations; spending limits and transaction observability across four AWS regions.
• OutSystems Agentic Systems Platform - 🆕 June 2026. OutSystems positions its low-code platform as an "AI-native" agentic development environment. Open and governed AI, BYO model, multi-agent orchestration, and enterprise compliance tooling. Challenger to Copilot Studio and Agentforce.
• Databricks Genie One - 🆕 June 16, 2026 (Data + AI Summit). Agentic "data coworker" that automates and orchestrates work across structured and unstructured data — inside or outside Databricks — grounded in the new Genie Ontology (an organization-wide knowledge graph) and governed by Unity Catalog. Ships with Genie Agents; Databricks reports 84.5% first-attempt accuracy in internal testing.
• ZenseAI.AgentMesh (Zensar) - 🆕 June 19, 2026. Zensar's enterprise-grade agentic AI platform — a "universal enterprise operating system for agentic AI" to discover, build, deploy, and govern autonomous agents at scale, with a catalogue of 80+ pre-built industry and cross-functional agents and a stated 6–8 week pilot-to-production path.
• Meta Business Agent - 🆕 June 3, 2026 (global rollout). Meta's AI business agent for WhatsApp, Instagram, and Messenger — answers customer questions, recommends catalog products, books appointments, qualifies leads, and closes sales, handing off to a human when needed. 1M+ businesses already on it; a Meta Business Agent Platform lets enterprises configure agents and connect Shopify / Zendesk / Shopee — free to activate today, paid subscription tiers coming.
• Snyk Evo Agentic Development Security (ADS) - 🆕 June 2026. Security and governance platform built specifically for autonomous AI coding agents; governs what agents use, what they do, and the code they generate in real-time.
• Cognizant Neuro AI + ServiceNow AI Agent - 🆕 June 2026. Cross-platform enterprise orchestration enabling ServiceNow agents to work natively within Cognizant's Multi-Agent Accelerator.
• Talkdesk Agent Builder - 🆕 June 2026. Low-code builder allowing business users to deploy production-grade AI agents in the contact center in hours instead of weeks.
• HelloTwin Digital Authority - 🆕 June 2026. AI twin designed as a single, auditable source of truth to govern agentic workflows with clear boundaries.
• Hellomatik - 💰 ⚠️ Freemium / Unverified. AI agent platform that turns a company's knowledge base into agents that answer, sell, and book across WhatsApp, email, and web. Integrations: Shopify, Stripe, Sage. Claimed 25–30% chat-to-sale conversion.
• OpenAI Presence - 🆕 July 22, 2026. OpenAI's enterprise agent deployment platform — enables businesses to deploy AI agents at scale for customer service and operations. Claimed 75% of phone support interactions handled without human escalation.

📊 Agent Evaluation & Observability

Tools for testing, evaluating, and monitoring AI agents in production.

• AgentBench - Multi-dimensional benchmark for evaluating LLMs as agents.
• PerspectiveGap - 🆕 2026 (arXiv 2606.08878). First benchmark for multi-agent orchestration prompt writing — 110 scenarios across 10 communication topologies (chain, star, tree, mesh, etc.), testing whether LLMs can compose prompts that make sub-agents coordinate effectively. Measures role-fragment assignment, information leakage rate, and topology-aware prompt design. Key finding: average combined pass rate only 17.2% across 33 models; GPT-5.5 leads at 62.0%. Opus 4.8 shows surprising weakness in orchestration prompting despite strong coding scores. MIT-licensed benchmark data + eval scripts; merged into OpenCompass and Inspect Evals (role-fragment + free-form prompt writing tasks, June 2026).
• ClawBench - 🆕 Live-web benchmark for browser and computer-use agents — 153 everyday online tasks across 144 real platforms in 15 categories (purchases, appointments, job applications), run against production websites rather than offline sandboxes. An interception layer captures and blocks the final submission request so nothing actually happens in the real world, then a two-stage scorer (HTTP interception → LLM judge) checks whether the agent submitted the right thing. Headline result: frontier models complete only a small fraction — Claude Sonnet 4.6 at 33.3%. Paper · Leaderboard
• LangSmith - Platform for debugging, testing, evaluating, and monitoring LLM applications.
• Helicone - Open-source LLM observability platform — logs, metrics, and traces.
• Braintrust - Enterprise-grade stack for building AI products — evals, prompt playground, logging. SDKs: braintrust-sdk-javascript and braintrust-sdk-python (repo split/renamed from braintrust-sdk).
• Arize Phoenix - AI observability & evaluation — traces, evals, and datasets. July 7, 2026: Metric Charts, Trace Search, expanded REST API; July 1 added granular LLM + tool span attributes via new OpenInference span attrs.
• Langfuse - Open-source LLM engineering platform — traces, evals, prompt management. Acquired by ClickHouse Jan 2026; March 2026 shift to an observations-centric data model, April 2026 added Langfuse Cloud Japan + Experiments + Langfuse Academy + LLM-as-a-Judge API; v4 self-host release queued.
• OpenLLMetry - Open-source observability for LLM applications based on OpenTelemetry.
• Weights & Biases Weave - Toolkit for developing, evaluating, and monitoring AI applications.
• SWE-bench - Benchmark for evaluating LLMs on real-world software engineering problems.
• Terminal-Bench - 🆕 Benchmark for terminal-based coding agent evaluation. Maintained by Harbor Framework.
• Harbor - 🆕 Framework for evaluating and optimizing agents and LLMs at scale — run Terminal-Bench 2.x and custom benchmarks across thousands of cloud sandboxes, generate RL rollouts; a Stanford × Laude Institute collaboration. Apache-2.0.
• Arena (formerly LMArena / LMSYS Chatbot Arena) - Crowdsourced AI benchmark using human preference voting; leaderboards now cover LLMs, image generation, and code models. LMSYS → LMArena (2025) → Arena (2026).
• Patronus AI - 💰 LLM evaluation/red-teaming company now positioned as a frontier research lab building digital world models and simulation infrastructure for agent training ($50M Series B); research artifacts include Lynx, FinanceBench, and GLIDER.
• DeepEval - Pytest-style LLM eval framework with 14+ built-in metrics (G-Eval, hallucination, faithfulness). Most-starred open-source eval lib in 2026. Apache-2.0.
• Agenta - 🆕 Open-source LLMOps platform combining prompt playground, prompt management, evaluation, and observability.
• LangSmith SDK - Official client SDK for LangChain's hosted observability platform.
• AutoEvals - Standalone library of best-practice LLM eval scorers (factuality, JSON validity, semantic similarity, etc.) by Braintrust. Drop-in for any framework.
• BenchClaw - ⚠️ Unverified. Self-described multi-dimensional agent evaluation harness (17-judge tribunal, deception detectors, 10 scoring dimensions). Repo is single-maintainer with very low independent adoption; the same submission was sent to 8+ awesome lists in parallel — one was merged at eudk/awesome-ai-tools, the rest are pending or declined. Listed for visibility, evaluate before relying on its scores.
• PromptEden - ⚠️ Unverified. Commercial AI-visibility monitoring service — tracks how ChatGPT, Claude, Gemini, Perplexity, Copilot, and Grok describe brands and which competitors they recommend, refreshed daily across 9+ platforms. Submitted to 10 awesome lists on the same day — promising category but listed for visibility only, evaluate before purchasing.
• Laminar - 🆕 Open-source observability platform purpose-built for long-running AI agents (Apache-2.0, YC S24). OpenTelemetry-native, transcript view, Signals, SQL over traces, browser-agent session replay.
• LangSmith Engine - 🆕 May 2026 (Interrupt 2026). Autonomous failure-diagnosis layer for LangSmith — clusters production failures into prioritised issues, root-causes them across traces and code, and proposes fixes for human review. Companion to the new SmithDB (Rust + DataFusion-backed agent observability database).
• AgentSight - Zero-instrumentation eBPF observability for LLM/coding agents. Captures syscall-level traces (process, file, network) without modifying the agent, enabling full-stack behavioral analysis. MIT.
• Prismix - Real-time status monitoring for 77+ AI services (OpenAI, Anthropic, Cursor, etc.) with status badges and API, AI news aggregator from 71+ sources, and an MCP server directory with 80+ servers. Free, no signup.
• Ceros (by Beyond Identity) - 🆕 June 16, 2026. Agentic AI trust layer for unified identity, observability, and governance — discovery/inventory, runtime policy enforcement, audit trails. (Unrelated to the interactive-content company of the same name.)
• Zoom Agent Performance Suite - 🆕 June 2026. Dedicated suite to test, validate, and optimize autonomous agent performance in customer-facing scenarios.
• AgentOps - 🆕 Agent monitoring, compliance, and testing toolkit with session replays; MCP server for zero-config observability of any MCP-connected agent; 5K+ stars.
• OpenTelemetry GenAI Semantic Conventions - Standardized spans, metrics, and events for GenAI clients, agents, tool invocations, and MCP — vendor-neutral tracing across any OTel backend (Arize, Langfuse, Helicone, Jaeger, etc.). Moved out of the core semconv repo into a dedicated GenAI repository.
• Tracecat - 🆕 Open-source security-automation platform that captures complete agent traces for SOC workflows — integrates AI agents with detection, enrichment, and response pipelines. AGPL-3.0 (with enterprise-edition exceptions).
• Langfuse v4 - 🆕 ⚡ v4.0.0, July 29, 2026. Major release for the self-hostable LLM observability stack: full-text search across inputs/outputs/metadata, a new filter search bar, monitors & alerts, and rebuilt Observations API v2 / Metrics API v2 that the project claims are up to 165× faster.
• AcruxCore - ⚠️ Unverified (new repo, single maintainer, no third-party adoption yet). Self-hosted or SaaS LLM-ops platform — prompt versioning, an AI gateway, tracing, a tool catalog, and evaluation runs. Apache-2.0.

🔬 AI Research Tools

Tools and platforms for AI/ML research, experimentation, and development.

• Hugging Face - The AI community's platform — models, datasets, and Spaces for ML research.
• Hugging Face Transformers - The de-facto standard library for models and training tooling.
• vLLM - 🆕 High-throughput LLM serving engine with PagedAttention.
• Ollama - Run LLMs locally with a simple API. Supports Llama, Mistral, Qwen, and more.
• LM Studio - Desktop app for running local LLMs with a user-friendly interface.
• SGLang - 🆕 Fast serving framework for large language and vision models.
• llama.cpp - LLM inference in C/C++ — run models on consumer hardware.
• MLX - 🆕 Apple's array framework for ML on Apple silicon.
• Unsloth - 🆕 Fine-tune LLMs 2x faster with 70% less memory.
• OpenRouter - Unified API for accessing 400+ AI models from 70+ providers.
• Weights & Biases - ML experiment tracking, dataset versioning, and model management.
• Label Studio - Multi-type data labeling and annotation tool.
• SmithDB - 🆕 May 2026 (Interrupt 2026). LangChain's purpose-built agent observability database. Rust on top of Apache DataFusion + Vortex, with object-storage backing for trace data — designed for the volumes and access patterns of agent traces.
• Strands Evals (AWS) - 🆕 AWS's evaluation framework for agent workflows — Case/Experiment/Evaluator structure with LLM-as-judge support; companion to Strands Agents SDK. Apache-2.0.
• DSPy - Programming — not prompting — language models; widely used in research for systematically optimizing prompts and pipelines.
• Elicit - 🆕 AI research assistant for literature and systematic review over a very large academic-paper corpus. July 15, 2026: shipped a public API and MCP server so agents and workflows can call its search and review capabilities directly. July 17, 2026: published a paper-search evaluation reporting it outperformed five other search systems on BioASQ (vendor-run evaluation).

📚 Learning Resources

Papers, courses, tutorials, and guides for understanding and building AI agents.

Papers

• ReAct: Synergizing Reasoning and Acting in Language Models - The foundational paper on reasoning + acting in LLMs.
• Toolformer: Language Models Can Teach Themselves to Use Tools - Teaching LLMs to use external tools autonomously.
• Generative Agents: Interactive Simulacra of Human Behavior - Stanford's generative agent architecture with memory and reflection.
• A Survey on Large Language Model based Autonomous Agents - Comprehensive survey of LLM-based autonomous agents.
• The Rise and Potential of Large Language Model Based Agents - In-depth analysis of LLM agent capabilities and future directions.
• Agent Hospital - A simulacrum of hospital with evolvable medical agents.
• ComBodied Agents: a New Paradigm of Human-Centric Agentic AI - 🆕 August 11, 2026. Human-centric agent paradigm combining multimodal perception, longitudinal memory, and personal world models that track human-state trajectories. Top Hugging Face daily paper on Aug 12, 2026.
• Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design - 🆕 August 10, 2026. Survey of agent-agent / agent-environment co-evolution and self-directed evolutionary mechanisms in agentic systems.

Courses & Tutorials

• DeepLearning.AI — AI Agents in LangGraph - Short course on building agents with LangGraph.
• DeepLearning.AI — Multi AI Agent Systems with crewAI - Course on building multi-agent systems.
• DeepLearning.AI — A2A Protocol - 🆕 Free course on Google's Agent-to-Agent protocol.
• LangChain Academy - Free courses on LangChain, LangGraph, and agent development.
• DeepLearning.AI Short Courses - Short AI courses covering the major frameworks / protocols.
• Hugging Face — Building AI Agents - Open course on building AI agents with open-source tools.
• LLM Agents MOOC (Berkeley) - UC Berkeley course on LLM agents (root site redirects to the latest iteration).
• Microsoft Agent Framework Docs - 🆕 Official documentation for Microsoft's unified agent framework.
• Hugging Face Agents Course - Free 5-unit course (notebooks + videos) on building production agents with smolagents, LangGraph, and Llama-Index.
• Anthropic Cookbook - Official notebooks for tool use, computer use, agent patterns, prompt engineering, and Claude Code recipes.
• Google Gemini Cookbook - Official Gemini API examples covering grounding, function calling, multimodal, and live audio.
• LLM Course (Maxime Labonne) - End-to-end LLM curriculum from fundamentals to fine-tuning, with Colab notebooks. 79K stars.
• Anthropic Courses - Anthropic's official educational courses on prompt engineering, real-world prompts, evals, and tool use.

Curated Lists

• awesome-ai-agents - 💤 Stale (last update 2025-02). Curated list of AI autonomous agents by E2B — pre-2026 reference.
• awesome-llm-agents - Curated list of LLM-powered agent resources.
• awesome-mcp-servers - 🆕 Curated list of MCP server implementations.
• awesome-ai-agent-papers (VoltAgent) - 🆕 Curated collection of 2026 AI-agent research papers — agent engineering, memory, evaluation, workflows, autonomous systems. Updated weekly from arXiv.
• awesome-cli-coding-agents - 🆕 Curated directory of terminal-native AI coding agents + the harnesses that orchestrate them — open-source tools (Pi, OpenCode, Aider, Goose), platform agents (Claude Code, Codex, Gemini CLI), parallel runners, autonomous loops.

🇨🇳 Chinese AI Ecosystem

Major projects from mainland-China teams or primarily targeting the Chinese market. Listed because the China stack is increasingly its own parallel ecosystem with distinct frameworks, models, and developer culture.

Foundation models from Chinese labs (Qwen, DeepSeek, GLM, Doubao, Kimi, Hunyuan, ERNIE) are listed under 🧠 Foundation Models directly.

Agent Platforms & Frameworks

• Dify - Open-source LLM app development platform with visual agent builder. The dominant low-code agent canvas in Chinese tech.
• LobeHub - Agent management platform (formerly Lobe Chat) — organizes agents into 7×24 operation with hiring/scheduling/reporting on your AI team. One of the highest-starred TypeScript AI projects (80K+ stars). Apache-2.0.
• Coze Loop - 🆕 ByteDance's open-source agent optimization platform from the Coze team.
• AgentScope - Alibaba ModelScope's multi-agent framework with visual debugging and distributed execution. Apache-2.0.
• Bisheng - Open enterprise LLM DevOps platform: workflows, RAG, agents, fine-tuning, evals. Apache-2.0.
• MetaGPT - Multi-agent collaboration framework that assigns SOP roles (PM, architect, engineer) to LLMs. Now maintained under the FoundationAgents org.

RAG / Knowledge

• FastGPT - Knowledge-base-first platform on top of LLMs: data ingestion, RAG retrieval, visual workflow orchestration.
• QAnything - 💤 NetEase Youdao's question-answering engine over arbitrary local documents (PDF/Word/Excel/PPT).
• RAGFlow - Deep-document-understanding RAG engine — strong on scanned PDFs, tables, and charts.
• LightRAG - HKU Data Science Lab's lightweight graph-based RAG engine.

Personal & Productivity

• AppFlowy - Open-source Notion alternative with AI workspace agents. AGPL-3.0.
• Manus AI - General-purpose autonomous agent by Butterfly Effect (Chinese-founded, relocated to Singapore). Meta announced a ~$2B acquisition on Dec 30, 2025, but China's NDRC blocked the takeover on April 27, 2026. ✅ Resolved August 11, 2026: Manus announced it will resume operating as an independent company as Meta unwinds the acquisition per Beijing's order; a user data-deletion process has begun.
• Coze (扣子) - ByteDance's no-code agent builder. Mainland-only consumer surface; international counterpart is coze.com.
• Qwen App (千问) - Alibaba's mass-market consumer agent (rebranded from Tongyi Qianwen), integrated across Taobao / DingTalk / Quark.
• Doubao Agents - ByteDance's flagship consumer assistant on top of the Doubao model family.

Developer Tools

• Trae - ByteDance's AI IDE and "10x AI coding engineer" — the highest-profile Chinese challenger to Cursor.
• CoderPlan - China-first unified LLM API gateway (Claude / OpenAI / Gemini, one-line config for Claude Code). Pay-as-you-go with Alipay & WeChat Pay.
• Cherry Studio - Most-installed open-source desktop client for LLMs in Chinese dev circles — multi-provider chat with knowledge base.
• Kimi Code CLI - 🆕 June 6, 2026. Moonshot AI's terminal coding agent (MIT, TypeScript) — built-in coder / explore / plan sub-agents in isolated contexts, conversational MCP setup.
• Qwen Code - Alibaba Qwen team's open-source terminal coding agent — agent teams, auto-memory, IDE integrations, multi-provider (OpenAI / Anthropic / Gemini / Qwen). 26K+ stars.
• Coze Studio - 🆕 ByteDance's open-source counterpart to Coze.com — all-in-one visual agent builder with debugging and deployment tools. Apache-2.0, 20K+ stars.
• ScienceOne 100 / 磐石100 - 🆕 Chinese Academy of Sciences scientific reasoning agent system, 50+ CAS institutes, 2,000+ research tools.

Notable 2026 Models (Chinese Labs)

• Kimi K3 - 🆕 🇨🇳 July 16, 2026. Moonshot AI's flagship MoE model — 2.8T parameters, 1M context window, Arena leaderboard front-end #1 (tied Fable 5), Artificial Analysis overall #3. Open weights released July 27, 2026. API: kimi-k3.
• Qwen 3.8-Max - 🆕 ⚡ 🇨🇳 Fully launched August 3, 2026 (previewed July 19 at the World AI Conference). Alibaba's next-generation MoE model — 2.4T total / 95B active parameters, 1M-token context, multimodal input; API on Alibaba Cloud Model Studio. Open weights on Hugging Face (Qwen/Qwen3.8-2.4T-A95B, ~Aug 12), with the Apache-2.0 multimodal 27B distillation following August 14.
• Tesla China OTA 2026.14.13 — Doubao Integration - 🆕 ⚡ 2026-07-31. Tesla pushes OTA to Model 3/Y/S/X in China replacing the original voice assistant with ByteDance's Doubao large model — first major Western EV brand to deep-integrate a Chinese LLM in-vehicle. ⚠️ Unverified (cnevpost URL not directly fetched; multi-source confirmed).
• DeepSeek V4 - 🆕 🇨🇳 July 2026. DeepSeek's latest flagship — 1M token context window, enhanced agentic task support, peak/off-peak tiered pricing. Continues DeepSeek's cost-efficiency positioning.

📝 Compare — Side-by-Side Tables

Quick decision matrices for the most common "which one do I pick?" questions in 2026.

🏗️ Agent Frameworks

🧪 Sandboxes (running agent-generated code)

🌐 Browser-Use Stacks

📊 Eval & Observability

💻 Coding Agents — Headline Picks

Tables verified 2026-07-17. Send PRs with sources when figures change.

💰 Foundation Models — API Cost & Context

Prices in USD per 1M tokens. Re-verified against official pricing pages on 2026-08-15 — see source notes below the table.

Source notes (2026-08-15 verification): - Anthropic figures from platform.claude.com/docs/en/about-claude/pricing and .../models/overview. Sonnet 5's introductory $2/$10 "is now the standard price — the previously scheduled increase to $3/$15 on September 1, 2026 will not occur." Haiku 4.5 is a 200K-context model, not 1M — an earlier revision of this table said 1M. - OpenAI figures from developers.openai.com/api/docs/pricing.md; context is documented as 1,050,000 tokens. Prompts above 272K tokens are billed at 2× input / 1.5× output (Terra long-context: $4.00/$18.00; Luna: $0.40/$1.80). - Google figures from ai.google.dev/gemini-api/docs/pricing; Gemini 3.7 Flash intro pricing runs through December 31, 2026. - DeepSeek moves to peak/off-peak tiered pricing at 16:00 UTC on August 16, 2026 (peak 01:00–04:00 & 06:00–10:00 UTC; off-peak 50% lower); cache-hit input remains dramatically cheaper ($0.044 peak for V4-Pro, $0.014 peak for V4-Flash). - GPT-4o / GPT-4o-mini were dropped from this table — GPT-4o left ChatGPT in Feb 2026 and the GPT-5.6 tiers are cheaper per unit of capability. They remain available on the API for legacy integrations.

💻 Foundation Models — Local Deployment

Estimated VRAM at Q4_K_M quantization. Speed varies by hardware. Weight availability verified against the Hugging Face API on 2026-07-30 — the HF repo column is the exact path, because near-miss names (there is no gemma-4-27b) are a recurring source of broken instructions.

VRAM figures are the weights-resident floor at Q4 — add headroom for KV cache, which at 1M context dominates the budget. For MoE models only the active experts need to be resident for compute, but the full weight set still has to be loaded or streamed, so treat the MoE rows as "fits with offloading", not "fits in 4 GB".

⚠️ Licensing is not the same as "open source." Kimi K3 ships under a bespoke Kimi K3 License with a revenue-threshold carve-out for managed-service providers, and Gemma models are under the Gemma Terms of Use — neither is OSI-approved. Apache-2.0 (Qwen, Mistral, Inkling) and MIT (Phi-4, DeepSeek V4) are the genuinely unencumbered options.

🧠 Agent Memory Systems

🎙️ Voice & Audio Models

🎨 Image Generation Models

🎥 Video Generation Models

🔍 RAG Frameworks

🗄️ Vector Databases

📱 Personal AI Assistants (2026)

🔌 MCP Servers — Top Integrations

Stars data approximate, 2026-05.

Use mcp-scan (Invariant Labs) to audit any MCP server before production deployment.

🏢 Enterprise AI Agent Platforms

📏 Embedding Models

MTEB = Massive Text Embedding Benchmark leaderboard score (EN, 2026-05 approx).

💡 The strongest open embedding option is now Apache-2.0 and self-hostable. The Qwen3-Embedding family tops multilingual MTEB while running locally — -8B for quality, -0.6B (10M+ downloads) when you need cheap bulk indexing. If you're still defaulting to text-embedding-3-large for a self-hosted RAG stack, benchmark against these first. There is no OpenAI text-embedding-4 as of 2026-07-30 — don't plan around one.

🛡️ Agent Security Tools

🖥️ Computer Use & Desktop Agents

🤖 Physical AI Platforms

🇨🇳 Chinese AI Models — Head-to-Head

Chinese language capability benchmarks are approximate. API prices in USD/1M tokens, July 2026.

📦 Agent Frameworks — TypeScript / JavaScript

📊 Meta-Comparison — Orchestration vs Framework vs IDE

📱 Mobile AI Frameworks

All Compare tables data: 2026-07-17. Send PRs with sources when figures change.

🗺️ Scenario Guide — What Should I Use For…

50+ curated scenarios matching your goal to the right tool or stack. Updated weekly.

🏗️ Building: Coding Agents

I want to build a coding agent for my startup (lowest cost, high quality) → Claude Code (CLI) + E2B sandbox + Langfuse observability. SWE-bench 80.9%. ~$200/mo at moderate usage.

I want an enterprise coding agent with security controls - GitHub Copilot Enterprise — Deep GitHub integration, IP indemnity, SSO/SAML, SOC 2. → best if already on GitHub Enterprise - Cursor Business — Privacy mode, code never leaves your infra, admin dashboard. → best for teams needing IDE-first UX - Devin 3.0 (Cognition) — Fully autonomous PR-to-merge with re-planning. → best for hands-off long-horizon tasks

I want an open-source self-hosted coding agent (no vendor lock-in) - OpenHands (All-Hands-AI) — MIT, competitive SWE-bench, BYO model. → best if you need full control - Cline (VS Code ext) — BYO key, large community, free. → best for VS Code users - Aider — Git-aware CLI refactoring, excellent polyglot support. → best for terminal-based git workflows

I want a browser automation / web scraping agent - Browser Use — 109K stars, vision + DOM, MIT. → best for general web automation - Stagehand (Browserbase) — Typed act/extract/observe API, structured output. → best for reliability-critical scraping - Skyvern — Vision-first, handles dynamic pages without CSS selectors. → best for changing / heavily JS-rendered sites

I want a document processing / PDF analysis agent → LlamaIndex (document pipeline) + Gemini 3.1 Pro (1M context, native PDF) or Claude Opus 5 (1M context, best reasoning) + Unstructured.io for ingestion. For local: Ollama + Qwen3.6-27B.

I want a customer service / support agent - Dify — No-code LLM workflow builder, self-hostable, RAG built-in. → best for non-technical teams - LangGraph + Zendesk MCP — Stateful workflows, ticket resolution loop. → best for engineering-led teams - Salesforce Agentforce — CRM-native, works within existing Salesforce data. → best for Salesforce-first orgs

I want a research / deep-research agent → Perplexity Deep Research (managed) or OpenHands + Tavily Search + Claude Opus 5. For local: Khoj (self-hosted). Expect multi-minute runs and $1–5 per deep report at cloud rates.

I want a data analysis / BI agent - Julius AI / Code Interpreter (ChatGPT) — Managed, no setup. → best for analysts without eng support - AI for Database — ⚠️ Unverified. Plain-English queries over Postgres / MySQL / MongoDB / SQL Server / SQLite + Sheets, with self-refreshing dashboards and Slack/webhook/email triggers; SOC 2 + GDPR, self-host option, $19/mo Pro. → best for non-technical teams that need direct DB access without SQL - LangChain + Pandas Agent + Langfuse — Fully custom, code-gen for queries. → best for eng teams with custom data - Metabase AI / Tableau Pulse — Embedded BI copilot. → best inside existing BI stack

I want a computer use / desktop automation agent - Claude Desktop Intelligence (Anthropic) — Screen-aware, controls any GUI app. → best all-around for macOS/Linux - UFO (Microsoft, open-source) — Windows-native, Win32 + UI Automation APIs. → best for Windows automation - Screenpipe — Continuous screen + audio recording + local LLM inference. → best for local privacy-first

I want a voice / conversational agent - Gemini Live API — Real-time voice, <300ms latency, Google cloud. → best for Google ecosystem - OpenAI Realtime API (GPT-Realtime-2 / GPT-Live-1) — Native voice with tool calling. → best for OpenAI ecosystem - LiveKit + Whisper + ElevenLabs v3 — Self-hostable voice pipeline. → best for custom, brand-specific voice

I want a multi-agent orchestration system - LangGraph — Stateful graph workflows, best Python production option. → best for complex state machines - OpenAI Swarm / Agents SDK — Lightweight handoffs, OpenAI-native. → best for simple OpenAI agent networks - Google ADK — Hierarchical agent coordination, Gemini-native. → best for Google/Vertex stack - Mastra (TypeScript) — Type-safe workflows, TS-first teams. → best for TypeScript stacks

I want a personal AI assistant (self-hosted) - OpenClaw — Multi-channel (Telegram/Discord/WhatsApp), memory, cron, MCP support, full local LLM option. → best all-in-one self-hosted - Khoj — Search + research + calendar, open-source, self-host. → best for knowledge workers - Jan.ai / LM Studio — GUI-first local model runners. → best for non-technical local LLM

I want a personal AI assistant (managed / easy setup) - Claude.ai (Pro) — Projects, memory, MCP tools, best reasoning. → best for power users - Perplexity Pro — Search-first, cites sources. → best for research-heavy use - ChatGPT Plus — Code interpreter, image gen, broad tools. → best for general-purpose

I want to build a RAG application → LlamaIndex (orchestration) + Qdrant (vector DB) + Cohere embed-v4 (embeddings) + BGE reranker (reranking). Managed alternative: Ragie or Cognee. Production telemetry: Langfuse.

I want a financial analysis agent → LangGraph + yfinance / Alpha Vantage MCP + Claude Sonnet 5 (Excel/table reasoning) + Langfuse. Avoid: don’t use hallucination-prone models for numbers — always validate with structured output + code execution.

I want a legal document agent → Claude Opus 5 (1M context, best contract analysis) + Claude for Legal connectors if you're already on Cowork + LlamaIndex (ingestion) + pgvector (self-hosted vector). Important: always have a human-in-the-loop for final legal decisions.

I want an education / tutoring agent - Khanmigo (Khan Academy) — Purpose-built for K–12, COPPA-compliant. → best for K–12 safe deployment - Custom with GPT-5.6 Terra + LangGraph state machine + spaced repetition logic. → best for HEd or corporate training

I want a creative writing assistant → Claude Opus 5 (best prose quality) or Gemini 3.1 Pro (long-form, 1M context) + Notion / Obsidian MCP for knowledge base. For structured fiction: Sudowrite (managed).

I want an IoT / physical AI agent → ROS 2 (robot OS) + NVIDIA Isaac GR00T (humanoid foundation model) + Genesis Sim (simulation). For home automation: Home Assistant + custom LLM backend.

I want a game playing / simulation agent → PettingZoo (multi-agent RL env) + Gymnasium + Gemini 3.5 Flash vision for game-state parsing. For LLM-in-the-loop games: Concordia (Google DeepMind).

I want a security scanning / vulnerability agent → Semgrep (static analysis) + Claude Sonnet 5 (explain + triage findings) + mcp-scan (MCP server audit). Managed option: Claude Security (Opus-powered codebase scanner). See also: Agent Security table.

I want a healthcare AI tool (non-clinical / administrative) → Claude Opus 5 + RAG on medical knowledge base + strict output validation. Always: disclose AI, human oversight for any clinical decision, check HIPAA/GDPR compliance. Never automate clinical diagnosis.

I want a code review / PR security agent → CodeRabbit (managed, instant PR reviews) or Claude Code in CI + Semgrep + custom rules. For enterprise: Copilot Code Review (GitHub).

I want a social media / content creation agent → n8n (workflow automation) + Claude Sonnet 5 (drafting) + gpt-image-2 (images) + Buffer/Later MCP (scheduling). Self-hosted option: all via n8n + Ollama.

I want a translation / localization agent → DeepL API (best quality for EU languages) or Claude Sonnet 5 (nuanced context-aware) + Weblate (open-source TMS). For Chinese: Qwen3.7-Max or Kimi K3 + human review loop.

🧠 Model Selection

I need the smartest model for complex multi-step reasoning - Claude Opus 5 — Anthropic's current flagship, near-Fable-5 quality at $5/$25 per M tokens. 1M context, 128K output. Default on Claude Max. - Claude Fable 5 — Mythos-class tier above Opus, for the genuinely hardest work. $10/$50 per M tokens. - GPT-5.6 Sol — OpenAI's frontier tier with "max" reasoning + "ultra" sub-agent mode. $5/$30 per M tokens. - Gemini 3.1 Pro — Google's shipped flagship, GPQA-Diamond record holder, 1M context. $2/$12 per M tokens. (Gemini 3.5 Pro with its reported 2M window is still unreleased — don't plan around it.)

I need the fastest + cheapest model for simple, high-volume tasks - Gemini 3.5 Flash-Lite — cheapest Gemini tier as of July 21, 2026; beats earlier Flash-Lite generations on high-throughput execution. (gemini-3.1-flash-lite remains GA if you're already on it.) - DeepSeek V4-Flash — $0.14/$0.28 per M tokens, 1M context, MIT. Best quality-per-dollar on the open-weight side. - Claude Haiku 4.5 — $1/$5 per M tokens, 200K context, Anthropic ecosystem integration. - GPT-5.6 Luna — $0.20/$1.20 per M tokens, fastest GPT-5.6 tier, broad OpenAI tooling.

I need the best Chinese language support - Qwen3.7-Max (Alibaba) — current Qwen flagship, purpose-built for agentic coding and long-horizon missions. → cloud API - Kimi K3 (Moonshot) — 2.8T total / 104B activated open weights (July 27, 2026), the largest openly available model at release. → self-host or API - DeepSeek V4-Pro — 1.6T/49B MoE, 1M context, MIT. → self-host or API - GLM-5.2 (Zhipu AI) — 1M context, MIT open weights, Chinese-first. → API or local

I need the best local/offline model with ~16GB VRAM - Qwen3.6-35B-A3B — MoE with only 3B active params, so it runs where a 35B dense model never would. Apache-2.0, best quality-per-GB pick. - Gemma 4 12B (Google) — Unified encoder-free multimodal (text + image + audio in one pass), ~7GB at Q4. - Phi-4 (Microsoft, MIT) — ~9GB at Q4, punches above its weight on coding. - Qwen3.6-27B — ~16GB at Q4; the strongest thing that still fits one consumer card.

I need the best local/offline model with 40GB+ VRAM - Qwen3 235B A22B — MoE flagship, 22B active, best local quality in this class. Apache-2.0. - Llama 3.3 70B Q4_K_M — ~42GB, strong English + coding, Llama Community License. - DeepSeek V4-Flash — 284B MoE / 13B active, MIT, 1M context, ~75GB at Q4 on a single 8×GPU node. - Inkling (Thinking Machines) — 975B MoE / 41B active, Apache-2.0, natively multimodal; needs multi-GPU.

I need the best coding capability → Claude Sonnet 5 (the most agentic Sonnet yet, default in Claude Code) for day-to-day agentic coding; Claude Opus 5 or Fable 5 when the task is genuinely hard. GPT-5.6 Sol via Codex for the OpenAI stack. DeepSeek V4-Pro for open-weight coding. For IDE use: Cursor (Grok 4.6 default since Aug 12, 2026; Claude backends available) or Cline.

I need multimodal understanding (vision + text) - Gemini 3.1 Pro — Native vision, PDF, audio, video understanding. 1M context. - GPT-5.6 Sol / Terra — Mature vision API, strong diagram/chart understanding. - Claude Opus 5 — Best for complex document image reasoning. - Inkling (Thinking Machines) — Apache-2.0, natively multimodal across text/image/audio/video in one model. - Qwen3-VL 72B — Best open-weight multimodal for self-hosting.

I need very long context (500K+ tokens) - Claude Opus 5 / Sonnet 5 — 1M context, best quality within that window. - Gemini 3.1 Pro — 1M context, cheapest strong option at this length. - DeepSeek V4-Pro / V4-Flash — 1M context, MIT, self-hostable. - Grok 4.5 — 500K context with context compaction built in. - Kimi K2.6 — 262K context, strong Chinese.

I need real-time voice / audio model - Gemini Live API — <300ms latency, native Google cloud. - OpenAI Realtime API — GPT-Realtime-2 (May 2026), native function calling during voice. - ElevenLabs v3 — Best TTS quality, 70+ languages. - Voxtral (Mistral) — Open-weight audio model, transcription + understanding.

I need the best image generation - gpt-image-2 (OpenAI) — Best instruction-following, 2K/4K, $0.04–0.17/image. - Flux 2 Pro (Black Forest Labs) — Photorealistic, fast, API available. - Midjourney V8 — Best artistic quality, no API (web only). - Stable Diffusion 3.5 — Open weights, local deployment, Apache-2.0.

I need the best video generation - Veo 3.1 (Google) — High fidelity, physics-aware, best quality 2026. - Kling VIDEO 3.0 (Kuaishou) — Leading post-Sora, strong cinematic style. - Runway Gen-4 — Precise motion control, professional use. - Seedance 2.0 (ByteDance) — Fast, cost-effective, strong for social media.

I need an open-weight model (MIT or Apache license) - Llama 3.3 70B (Meta, Llama 3.3 Community License) — Best English open-weight. - Qwen3 235B A22B (Alibaba, Apache-2.0) — Best Chinese + coding open-weight. - Mistral Small 4 (Mistral AI, Apache-2.0) — 119B MoE / 6B active, fast and multilingual. - DeepSeek V4-Flash / V4-Pro (MIT) — Best open-weight coding. - Inkling (Thinking Machines, Apache-2.0) — 975B MoE / 41B active, natively multimodal, 1M context. - Gemma 4 31B / 12B (Google, Gemma Terms of Use — not OSI) — Strong multilingual reasoning; 12B is the unified multimodal variant.

🏗️ Infrastructure

I want to run everything locally (privacy-first, zero cloud) → Ollama (model runner) + Open WebUI (UI) + Qdrant (local vector DB) + Qwen3.6-27B (16GB VRAM) or Llama 3.3 70B (40GB+). Full stack: OpenClaw (local mode) or AnythingLLM. No data leaves your machine.

I want to minimize API costs (budget <$50/month) → Use DeepSeek V4-Flash ($0.44/$1.32 peak, half off-peak, from Aug 16, 2026) or Gemini 3.5 Flash-Lite for high-volume. Reserve Claude Sonnet 5 for complex tasks only — its $2/$10 intro pricing was made permanent on Aug 10, 2026. Use Anthropic Batch API (50% off) for non-real-time work. Cache aggressively: DeepSeek cache hits cost a few percent of a cache miss, and Anthropic prompt caching reads at $0.50/MTok against $5.00 base input on Opus-class models. Schedule bulk DeepSeek jobs off-peak for a further 50% saving.

I want to scale to enterprise (millions of requests/month) → Google Vertex AI (managed Gemini, auto-scale, SLAs) or Azure OpenAI / Microsoft Foundry (GPT-5.6 tiers plus Grok 4.3, compliance, dedicated capacity). Add Langfuse for observability. For routing: PortKey or LiteLLM as unified gateway.

I want to deploy in an air-gapped / regulated environment → Ollama (local inference) + Qwen3 235B A22B / Llama 3.3 70B (open weights) + Qdrant (local vector DB). For enterprise needs: IBM watsonx (on-prem) or Azure Government (FedRAMP). Compliance certifications matter more than model quality here.

I want to build for edge / mobile deployment → Core ML (Apple, iOS/macOS) + Phi-4 14B or Gemma 4 2B (quantized). For Android: MediaPipe + Gemma 4. For cross-platform: llama.cpp + GGUF models. Check Qualcomm AI Hub for Snapdragon-optimized models.

I need multi-cloud / want to avoid vendor lock-in → LiteLLM (unified API proxy for 100+ providers) + LangGraph (framework-agnostic) + provider-agnostic embeddings (BGE-M3 open-weight). Store all state in self-hosted Qdrant or Postgres+pgvector.

I want to self-host everything (no managed services at all) → Ollama (models) + Qdrant (vectors) + Langfuse (observability, self-host Docker) + n8n (workflows) + OpenClaw (agent runtime). GPU recommendations: 2× RTX 4090 (24GB each = 48GB) for 70B models; single RTX 4090 for 27B.

📊 Evaluation & Monitoring

I want to evaluate agent output quality → DeepEval (rich metric suite: faithfulness, relevancy, hallucination) + Langfuse (traces + evals). For custom: LangSmith (tight LangChain integration). Open-source: Agenta (self-host).

I want to debug why my agent is failing → Langfuse (trace each tool call + LLM call with timing) + Arize Phoenix (root cause analysis). Enable verbose logging in your framework (LangGraph / CrewAI all support it). Use LLM-as-judge to flag low-confidence steps.

I want to monitor production agents in real-time → Langfuse (OpenTelemetry-native, self-host) or Helicone (zero-latency proxy logging). Set up cost + latency + error-rate dashboards. Alert on error spikes via Grafana or Datadog integration.

I want to A/B test different models or prompts → Braintrust (experiment tracking, online/offline eval) or LangSmith (prompt playground + evals). For open-source: Agenta + Langfuse experiments feature.

I want to benchmark models on my specific tasks → LMSYS Chatbot Arena custom eval + Evals by OpenAI (open framework) + DeepEval (custom metric). Run your own eval harness: prepare 50–200 golden examples, measure precision/recall on your actual task.

I want to evaluate MCP server security → mcp-scan (Invariant Labs) — Detects prompt injection, tool poisoning, shadow tools in MCP servers. Run before production deployment. See also: Agent Security.

🌍 Ecosystem Choices

I want to build within the OpenAI ecosystem → OpenAI Agents SDK + GPT-5.6 Terra + E2B sandbox + LangSmith eval. Benefits: widest third-party tooling, most community examples. Cost: premium pricing.

I want to build within the Anthropic Claude ecosystem → Claude Code (agentic IDE/CLI) + Claude Sonnet 5 / Opus 5 + MCP protocol (Claude Desktop) + Langfuse (observability). Benefits: best coding quality, MCP is Claude-native. Cost: ~mid-tier.

I want to build within the Google Gemini ecosystem → Google ADK v2.5+ (Agent Development Kit) + Gemini 3.1 Pro / 3.5 Flash + Vertex AI (deployment) + Vertex AI Eval + AlloyDB / BigQuery (data). Benefits: 1M context, strong multimodal, cheap Flash tier. Cost: scales well.

I want to build for the Chinese market (domestic cloud / regulation) → Qwen3 235B (Alibaba Cloud DashScope) + Baidu ERNIE 5 or Kimi K2.6 (Moonshot) as fallback + Alibaba Cloud PAI (deployment). All data stays within China borders. ICP-compliant.

I want a TypeScript-first stack → Mastra (TS workflows, MCP, A2A, Elastic-2.0) + Vercel AI SDK (streaming, RSC-friendly) + Qdrant JS client + Langfuse JS SDK. Alternative: LangChain.js + LangGraph.js.

I want an open-source-only stack (zero proprietary) → Ollama + Qwen3.6-27B or DeepSeek V4-Flash (model) + LangGraph (MIT, framework) + Qdrant (Apache-2.0, vector DB) + Langfuse (MIT, observability) + E2B (Apache-2.0, sandbox). Fully self-hosted, no vendor dependencies.

📋 Stack Recipes — Curated Tool Combinations

8 battle-tested multi-tool setups for common use cases. Copy and adapt.

⚠️ Anti-Picks — What NOT to Use For…

Avoid common mistakes. These recommendations are based on observed production failures in 2026.

🌟 Notable Agent Projects of 2026

Standout projects and developments that shaped the AI agent landscape in 2026.

• Model Context Protocol (MCP) - Became the universal standard for agent-tool interoperability. Donated to Linux Foundation.
• A2A Protocol - Agent-to-Agent protocol (originated at Google, now Linux Foundation) enabling cross-framework agent collaboration with 150+ partners; v1.0 shipped May 2026.
• Claude Code - Anthropic's agentic coding tool became the go-to terminal-based coding agent with 80.9% SWE-bench.
• Kiro - 🆕 AWS launched an autonomous coding agent capable of managing 10 simultaneous development tasks.
• Devin 3.0 - 🆕 Evolved to include dynamic re-planning, self-healing code, and legacy codebase migration.
• Microsoft Agent Framework - 🆕 AutoGen + Semantic Kernel merged into unified enterprise agent platform.
• OpenAI Codex CLI - OpenAI entered the agentic coding space with an open-source terminal agent.
• Browser Use - Breakthrough in making AI agents interact with the web naturally.
• Claude Computer Use - 🆕 Desktop Intelligence let Claude control any software by seeing the screen.
• Manus AI - General-purpose autonomous agent that can handle research, coding, and complex workflows. Meta's ~$2B December 2025 acquisition was blocked by China's NDRC on April 27, 2026 — the first time Beijing used its foreign-investment security review to stop an AI acquisition. On August 11, 2026 Manus announced it will resume operating as an independent company as Meta unwinds the deal.
• OpenHands - Open-source AI software engineering platform gained massive adoption.
• Dify - Low-code LLM agent platform reached mainstream adoption.
• Cline - VS Code autonomous coding agent with rapid community growth.
• Mem0 - Memory layer for AI became essential component of agent architectures.
• Sora Discontinuation - 🆕 OpenAI shut down Sora (April 2026), signaling strategic pivot to enterprise AI and reasoning.
• Kling VIDEO 3.0 - 🆕 Kuaishou's video generation became the leading AI video platform post-Sora.
• Cohere + Aleph Alpha Merger - 🆕 April 24, 2026. Canadian AI firm Cohere merged with Germany's Aleph Alpha at ~$20B valuation. $600M Series E from Schwarz Group. Creates transatlantic "sovereign AI" powerhouse with dual HQ in Toronto and Germany.
• ScienceOne 100 / 磐石100 - 🆕 April 28-29, 2026. Chinese Academy of Sciences launches specialized scientific AI system. 2,000+ research tools, 50+ CAS institutes. Flagship-level scientific reasoning and agent capabilities.
• Google Invests $40B in Anthropic - 🆕 April 2026. $10B initial + up to $30B contingent on performance milestones. Includes 5GW compute capacity over 5 years. Largest AI partnership investment to date.
• OpenAI Deployment Company (DeployCo) - 🆕 May 11, 2026. OpenAI spins out a $4B+ enterprise-deployment services unit (TPG / Bain Capital / Brookfield / Advent / Goldman Sachs / SoftBank + Bain & Company / Capgemini / McKinsey) and absorbs the Tomoro consulting acquisition. Signals the AI vendor race shifting toward services + Forward Deployed Engineers.
• Anthropic ↔ SpaceX Colossus 1 - 🆕 May 6, 2026. Anthropic takes all available capacity on the 300+ MW / 220K-GPU Colossus 1 Memphis cluster. SpaceX repositions itself as an AI infrastructure provider after its xAI acquisition; Anthropic doubles Claude Code rate limits for paid plans.
• DeepSeek $4B state-backed round - 🆕 May 16, 2026. China's National AI Industry Investment Fund + Big Fund III + Tencent close in on a ~$4B first external round for DeepSeek at a ~$50B valuation — first known LLM investment from Big Fund III, signalling Beijing's bet on efficient open-weight frontier models and domestic silicon.
• Pope Leo XIV → Vatican AI Commission - 🆕 May 16, 2026. Pope Leo XIV publishes the rescriptum establishing an inter-dicasterial Vatican commission on artificial intelligence (Dicastery for Integral Human Development coordinating, with Doctrine of the Faith, Culture & Education, Communication, Pontifical Academies for Life / Sciences / Social Sciences). One-year renewable mandate. First AI-focused encyclical expected to follow.
• Google I/O 2026 — Gemini 3.5 + Omni + Spark + AI Ultra - 🆕 May 19, 2026. Google's biggest agent-and-AGI keynote of the year: Gemini 3.5 Flash GA (default model), Gemini Omni world-model family, Gemini Spark 24/7 personal agent with ~30+ MCP-based tool integrations, and a new Google AI Ultra $100/mo tier. Pichai confirms Google now processes 3.2 quadrillion tokens / month.
• Alibaba Cloud Summit Hangzhou — Qwen 3.7-Max + Zhenwu M890 - 🆕 May 20, 2026. Alibaba unveils Qwen 3.7-Max (agentic-coding flagship for long-horizon missions), the T-Head Zhenwu M890 AI accelerator, and a full-stack AI infrastructure upgrade — China's most aggressive bid yet to position itself as the country's "AI factory."
• OpenAI Guaranteed Capacity (Compute Annual Pass) - 🆕 May 19, 2026. Long-term enterprise compute reservations (1/2/3-year terms) sold as a structured product — OpenAI's structural answer to Anthropic's Priority Tier and the wider supply crunch for frontier-model inference.
• JADEPUFFER — First Agentic Ransomware - 🆕 ⚠️ July 2, 2026 (Sysdig Disclosure). The first fully LLM-orchestrated ransomware operation: an AI agent exploited a Langflow RCE, self-corrected a failed step in 31 seconds, pivoted laterally, and encrypted 1,342 config items with an ephemeral key — making the ransom demand simultaneously unpayable and unrecoverable. Ransomware "narrated itself" with natural-language reasoning comments. Reference incident for why agentic security tooling is now considered production-critical.
• Kimi K3 Open Weights - 🆕 ⚡ Shipped July 27, 2026 — as promised. Moonshot fully opened the 2.8T-parameter (104B activated) Kimi K3 weights on Hugging Face, making it the largest openly available language model at time of release and the first Chinese frontier lab to open-source at 3T scale. Ungated, 97 safetensors shards, MXFP4/MXFP8 quantization-aware training; bespoke Kimi K3 License with a $20M-revenue MaaS carve-out.
• GPT-5.6 Sol National Security Delay - 🆕 July 9, 2026 landmark. GPT-5.6 launch was briefly delayed at US government request for national-security review — first time a government review directly delayed a major commercial AI model release. Rolled out in stages to trusted-partner preview, then general access following review clearance. Sets precedent for future frontier-model regulatory engagement.
• Robinhood Agentic Trading + Robinhood ↔ MCP - 🆕 May 27, 2026 (beta). First major US broker to open its trading API to AI agents via MCP. Read access to all accounts, trade-execute access only inside a ring-fenced Agentic account; push-notification on every trade and one-tap kill switch. Significant step in agentic finance — agents now hold real custody decisions, not just recommendations.
• Microsoft Scout + MAI-Code-1-Flash + MAI-Thinking-1 (Build 2026) - 🆕 June 2, 2026 (Build 2026). Microsoft simultaneously launches its first OpenClaw-based always-on personal agent (Scout), its first in-house coding model (MAI-Code-1-Flash, GitHub Copilot), and its first in-house reasoning model (MAI-Thinking-1). Together they mark Microsoft's biggest foundation-model independence move since the OpenAI partnership began.
• Meta Business Agent (WhatsApp + Instagram) - 🆕 June 3, 2026. Meta makes its customer-support AI agent globally available on WhatsApp + Instagram DMs at Conversations 2026 (London). Answers questions, recommends products, books appointments, qualifies leads; 1M+ businesses already using it. Tiered pricing tied to WhatsApp Business Premium — first Meta AI product Meta is monetising directly.
• WWDC 2026 — Apple Intelligence x Gemini + Foundation Models framework expansion - 🆕 June 8, 2026. Apple reveals the next-gen Apple Intelligence + the redesigned Siri AI (multimodal, screen-aware, on-device + server-routed), now powered by Google Gemini instead of the previous ChatGPT handoff. The Foundation Models framework expands to support image input, custom skills, and unified on-device + server models behind one Swift API; SiriKit deprecated in favor of expanded App Intents. EU/China launch deferred.
• Google Antigravity 2.0 + Microsoft RAMPART + xAI Grok Build - 🆕 May 14–22, 2026. Three structural agent-stack shifts in one week: Google's standalone multi-agent desktop + SDK at I/O 2026, Microsoft open-sourcing agentic-AI safety testing (RAMPART + Clarity), and xAI entering the CLI-agent race with Grok Build on grok-code-fast-1. Major / Anthropic-Google-Microsoft / xAI all show up with agent platforms within the same 8-day window.

📅 2026 AI Timeline

Key milestones and events in the AI landscape of 2026.

Contributing

Contributions welcome! Please read the contributing guidelines first.

License

This list is released under MIT License.