TOP 10 is unchanged for the third day; rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are identical to yesterday. Claude Fable 5.2 (65) leads, followed by GPT-6.2 Astra (62), Gemini 4 Neon (61), Gemini 4 Carbon (60), Claude Opus 5.5 (59), GPT-6.2 Sol (57), Claude Sonnet 5.5 (55), Gemini 4 Hydrogen (54), GPT-6.1 Terra (51) and Gemini 4 Argon (50). Google has four models on the list, while Anthropic and OpenAI have three each. Muse Wonder Full is tied with Argon at 50 points but remains outside the list because it ranks 11th in the table; Artificial Analysis still does not show an API provider for this model. The watchlist is also unchanged: Muse Wonder Full (50), Kimi K3.1 Pro (46, the strongest open-weight model), Muse Wonder (44), Qwen4 Max (43), Claude Haiku 5.5 and Muse Ember (36). The only new model to enter the table in the past 72 hours is Thinking Machines’ Apache 2.0-licensed Inkling 2 Mini (20 points). LMArena’s text table is available today, but its latest view is still dated November 11; Gemini 4 Carbon, Gemini 4 Argon and Muse Wonder lead the top three.
DeepSeek had the busiest update day: 75 entries came from the DeepSeek Harness v0.2.3-alpha.1 preview, released on the morning of November 22 (03:19 UTC). The plugin marketplace now accepts Git repositories as sources, including private repositories; optional connectors have arrived for Linear, Notion, Tencent Docs, Feishu, WeCom, Tencent Meeting and DingTalk, with connector accounts managed in one place using MCP OAuth and automatic credential refresh. A new code review page collects GitHub PRs relevant to you; the same plugin can review and merge GitLab merge requests. The experimental Timeline plugin shows the agent task tree and blockers; Computer Use is available experimentally on Windows and can be exposed to external tools as an MCP server with `cua-mcp`. The `/office` command combines Word, PowerPoint and Excel instructions, model settings now let you select a thinking level, and MCP tools can be loaded on demand to save context. This is an alpha release with breaking changes: Python PTC has been removed (use Node PTC), `enter_worktree`/`exit_worktree` replace the `working_directory` parameter, old module paths have been deleted, and plugin configuration has changed with Cordis 4.0.0-rc.12.
Eight of OpenAI’s 18 entries are from Ruby SDK 1.0: the new version requires Ruby 3.3+; the Assistants, Threads and Videos/Sora 2 surfaces and three old model constants have been removed, and compaction output has moved to the `ResponseItem` union. For Python, Agents SDK v1.0.0-alpha.1 introduces a preview of `CodexProvider`, which runs the same agent with Codex; Codex Core is embedded in the package with PyO3, and ready-made wheels are available only on GitHub Releases. Go SDK v3.105.0 and Ruby SDK v0.130.0 have gained Live WebRTC standby support. In ChatGPT, Business Agent publishing is now available to eligible e-commerce and online services businesses outside the US. At Anthropic, Claude Mobile’s new look is rolling out gradually on iOS 26 and later; Claude Data, announced on November 10, has become the default in Enterprise. Gemini can now generate looping GIFs of about three seconds with Gemini Omni 1.2 Flash (in the US, Japan, South Korea, Australia, Canada, the United Kingdom, Germany and France). In the US, users can place grocery orders from Albertsons, Safeway, Shaw’s and Target through Gemini and Search; connections to Experian and H&R Block are also coming to Search. Existing subscribers can get up to 50% off annual Google AI Pro and Google AI Plus plans through November 26. On the enterprise side, the Semantic governance policies preview checks every tool call passing through Agent Gateway against natural-language rules before execution, and also supports Anthropic, OpenAI, Mistral and vLLM endpoints. Agent Anomaly Detection can also monitor agents on AWS, Microsoft Foundry and on-premises in private preview.
Tencent’s 24 entries are from open models and LoopForge. The WeMM-Video family (2B, 4B, 9B and 35B-A3B) is available on Hugging Face as open-weight models that jointly process text, images, audio and video; it has instruct and thinking modes. In LoopForge, CodeBuddy is no longer required: descriptions, Bot, reports and scheduled tasks run with the selected CLI and model, and personal executor defaults have been added. Chat images are sent directly to six CLIs, scheduled AI tasks now have a 1–240-minute time limit and the `runs_wait` tool, and isolated sandboxes are now the default for new SOPs. The Access key screen provides step-by-step MCP setup for Cursor, Claude Code, CodeBuddy and Codex. Nine of Alibaba’s 11 entries come from the previously unrecorded portion of yesterday’s Qwen Code v0.28.3: an experimental WeCom channel adapter, preventing deletion of sessions with running background jobs, context-compaction recovery based on the model window, and a 10,000-character limit for async Hook output. Together AI announced that it offers Qwen4-Max on Serverless (not confirmed by Alibaba); five old Qwen cards on Hugging Face now redirect to newer versions. Meta’s FAIR team released ATLAS, a Lean library with more than 45,000 verified definitions and theorems from 26 open textbooks; DAT iOS 1.1.0 for Meta AI glasses adds session pausing, a MockDevice MCP server and clearer streaming errors. Z.ai is giving 50M tokens to users who log in to AutoClaw between November 23–29 and introduced the Legal assistant’s case-law research linked to official rulings. Xiaomi MiMo Code’s main branch now live-refreshes the skill catalog and adds `view=brief` to the session list API (no tagged release yet). Moonshot had only two Kimi Code Desktop Canary builds without release notes; they were not archived. xAI had no new changes between November 20 and 22. A total of 149 new developments were added to the archive today.