YZ·RADARyz.mikail.net
Daily edition · 22 November 2026

TOP 10 unchanged; 75 DeepSeek Harness updates, GIF generation in Gemini

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.

Today’s TOP 10

The ten best unique models by intelligence index; each model is counted at its highest-scoring setting.

  1. 01

    Claude Fable 5.2

    Anthropic · Claude=
    65pts

    Released on November 9; leads with 65 points, with its score and ranking unchanged. Available on all paid Claude plans; `claude-fable-5-2` on the API, $8 / $40, with thinking always adaptive. Claude Meetings and computer use betas are opening on Team and Enterprise plans; Claude Data has been enabled by default in Enterprise since November 20.Released: 9 November 2026

  2. 02

    GPT-6.2 Astra

    OpenAI · ChatGPT=
    62pts

    Released on November 17; remains in second place with 62 points. Task cost is $4,00 with the Max setting. `gpt-6.2-astra` on the API, $8 / $40, cached input $0,40; 1,05M context, 128K output. Sol is the default model in Codex; Astra must be selected manually from the model picker for difficult tasks. Also available in GitHub Copilot (Pro+, Max, Business, Enterprise).Released: 17 November 2026

  3. 03

    Gemini 4 Neon

    Google · Gemini=
    61pts

    Released on November 18; remains in third place with 61 points. Task cost is $6,44 with the xhigh setting. `gemini-4-neon` on the API (paid tier only), $3 / $15, 1M context; thinking levels include low, medium, high and xhigh. Available to Google AI Pro and Ultra subscribers in the Gemini app and Antigravity. Google is offering existing AI and storage subscribers up to 50% off annual Google AI Pro and Google AI Plus plans through November 26.Released: 18 November 2026

  4. 04

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    Its score (60) and ranking are unchanged; first in LMArena’s latest accessible view (November 11). `gemini-4-carbon` in Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, $4 / $20 from December 1).Released: 28 October 2026

  5. 05

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Its score (59) and ranking are unchanged. Anthropic’s flagship Opus model; available through Claude Pro/Max/Team/Enterprise and the API, as well as Bedrock, Google Cloud and Azure.Released: 22 September 2026

  6. 06

    GPT-6.2 Sol

    OpenAI · ChatGPT=
    57pts

    Released on November 17; remains in sixth place with 57 points. Task cost is $1,01 with the Max setting, a quarter of Astra’s. `gpt-6.2-sol` on the API, $2 / $10, cached input $0,10; 1,05M context, 128K output. The default model in the Codex model catalog. Also available in ChatGPT Work and GitHub Copilot.Released: 17 November 2026

  7. 07

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

    The strongest model available even on claude.ai’s free plan; its score (55) and ranking are unchanged.Released: 28 September 2026

  8. 08

    Gemini 4 Hydrogen

    Google · Gemini=
    54pts

    Released on November 12; remains in eighth place with 54 points. Task cost is $3,16 with the xhigh setting. `gemini-4-hydrogen` is also available on the free API tier; paid pricing is $1,50 / $7,50. Tasks in Gemini runs on Hydrogen; CodeMender supports Hydrogen in preview.Released: 12 November 2026

  9. 09

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

    Released on October 26 in ChatGPT Work, Codex and the API; focused on image and computer use. Its score (51) and ranking are unchanged. There is no GPT-6.2 version yet.Released: 26 October 2026

  10. 10

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available on Google AI Pro/Ultra and the paid API; promotional price of $2 / $10 through November 30. Tied with Muse Wonder Full at 50 points; retains tenth place because it ranks ahead of Muse Wonder Full in the Artificial Analysis table.Released: 30 September 2026

Key signals

  • TOP 10 unchanged: Claude Fable 5.2 (65) leads, with all scores and rankings identical to yesterday.
  • DeepSeek Harness v0.2.3-alpha.1: GitHub/GitLab code review, 7 new connectors, Computer Use on Windows.
  • Gemini generates looping GIFs with Omni 1.2 Flash; available in 8 countries, including France and Germany.
  • OpenAI Ruby SDK 1.0 is out: Ruby 3.3+ required; Assistants, Threads and Videos surfaces removed.
  • Tencent WeMM-Video is now open-weight; LoopForge no longer requires CodeBuddy.
  • 149 new developments: DeepSeek 75, Tencent 24, OpenAI 18, Alibaba 11, Meta 8, Google 7, Anthropic 2, Xiaomi 2, Z.ai 2.

Worth trying today

01

Try the DeepSeek Harness alpha in a separate workspace

v0.2.3-alpha.1 is a preview release with breaking changes. If you have your own plugin or SDK integration, consult the upgrade guide first: Python PTC has been removed, so you need to move your code to Node PTC; use `enter_worktree`/`exit_worktree` instead of `working_directory` and `cwd`; import types from `/types` or the package root; plugins that read session status must declare the `dsh-client-ui-session-status` package in `inject`. If you only use the interface, useful updates include connecting your GitHub account to the code review page and reviewing PRs with the agent, adding a Notion or Linear connector through OAuth on the Plugins page, enabling optional tool loading in MCP settings to reduce context usage, and lowering the thinking level in model settings for simple tasks.

DeepSeek · DeepSeek
02

Plan your Ruby SDK 1.0 migration on a separate branch

Don’t upgrade the `openai` gem to 1.0 immediately; first pin it to `~> 0.130` in your Gemfile and do the migration on a separate branch. Upgrade your Ruby version to 3.3 or later. If you use `client.beta.assistants` and `client.beta.threads`, create a Conversation for each thread and migrate to the Responses API; remove `client.videos` and Sora 2 calls; account for the `ResponseItem` union in code that handles compaction output and for the image `size` field now being a string. In voice applications, enable WebRTC standby with Ruby 0.130.0 and Go 3.105.0 (`session.wake_policy.type: speech`, `session.idle_timeout_s` 5–60 sec); a standby minute costs half as much as an active minute.

OpenAI · ChatGPT
03

Generate GIFs in Gemini and check AI plan discounts before November 26

In eight countries, including France, Germany and the United Kingdom, select GIF from the format selector in a Gemini chat or the Image/Video tool, then describe what you want; upload a single reference image for character or product consistency and describe the beginning and end of the loop. If you’re considering switching to Google AI Plus or annual Google AI Pro, the up-to-50% Black Friday offer for existing subscribers ends on November 26; verify the offer on the Google AI plans page and the renewal price at checkout. If you use enterprise agents, first run the Semantic governance policies preview in dry run mode and move to blocking only after reviewing the decisions in Cloud Logging.

Google · Gemini
04

Set executor defaults and task duration in LoopForge

If you use LoopForge, save the default CLI, specific model, reasoning effort and approval mode under Executors in your personal configuration; if CodeBuddy isn’t installed, you can now run Bot, reports and scheduled tasks with another executor. For scheduled AI tasks, set the time limit between 1–240 minutes to cover the SOP’s normal duration; jobs that hang are stopped when time runs out. To connect LoopForge MCP to Cursor, Claude Code or Codex, create an access key, copy the client-specific configuration shown on screen and run the verification step; don’t put the key in a chat or screenshot.

Tencent · Hy

This edition’s updates by AI

What changed in this edition for every AI in the TOP 10 and on the watchlist.

Anthropic · Claude

#1 Claude Fable 5.2#5 Claude Opus 5.5#7 Claude Sonnet 5.5Claude Haiku 5.5
Full history (836) →

New in this edition · 2

  • SettingClaude Data

    Claude Data is now enabled by default on Enterprise plans

    The Enterprise default announced at the November 10 general availability launch has taken effect: Claude Data is enabled by default starting November 20, 2026, for Enterprise organizations that have not previously saved a setting. It was already enabled on Team plans; Certified queries and live sharing beta features remain disabled by default on Enterprise.

    Usage details

    Who can use it: Enterprise plans; existing preferences are preserved if an owner has previously saved a setting. Claude Data is already enabled by default on Team plans.

    How to use it: In Claude, go to Organization settings > Artifacts > Templates; turn the Data template on or off to manage the organization default and access.

    Tip: Before enabling Data, check your organization’s data connectors and sharing permissions; Certified queries and live sharing settings are managed separately from Data.

  • FeatureClaude Mobile

    Claude Mobile’s refreshed look is rolling out gradually on iOS

    The refreshed interface for Claude Mobile is rolling out gradually, starting with Claude for iOS. Chats, projects, and plans remain unchanged; no separate setup is required.

    Usage details

    Who can use it: Claude for iOS users; iOS 26 or later is required for the new look, and the rollout is gradual. This announcement did not specify a separate date or scope for Android.

    How to use it: Update the Claude app from the App Store and upgrade to iOS 26 or later; the look appears automatically, so you don’t need to enable a setting.

    Tip: If the new look doesn’t appear, check that the app is up to date and the device is running iOS 26 or later.

    ✓ Verified by official sourceSource: Release notes | Claude Help Center

OpenAI · ChatGPT

#2 GPT-6.2 Astra#6 GPT-6.2 Sol#9 GPT-6.1 Terra
Full history (339) →

New in this edition · 18

  • APIOpenAI Ruby SDK

    OpenAI Ruby SDK 1.0 released and requires Ruby 3.3+

    The first 1.0 release of the OpenAI Ruby SDK is now available; the new version requires Ruby 3.3.0 or later and includes extensive API type and behavior changes. OpenAI has published a separate migration guide.

    Usage details

    Who can use it: Developers using the OpenAI Ruby SDK who can run Ruby 3.3+; version 1.0.0.

    How to use it: Upgrade the Gem dependency to 1.0.0, set the Ruby runtime to 3.3+, and test and apply the breaking changes described in the migration guide.

    Tip: Before upgrading production, run builds, type checks, and all Responses/Live/Videos flows on a separate branch.

  • APIOpenAI Ruby SDK

    Ruby SDK Responses compaction output switches to the ResponseItem union

    GA and beta compaction output now uses the broader `ResponseItem` union, which can include preserved user/system/developer messages and compaction records, not just assistant output. Compressed function call items should also use `ResponseFunctionToolCallItem` or its beta counterpart.

    Usage details

    Who can use it: Ruby applications that consume `responses.compact` or beta compaction output with type annotations.

    How to use it: Update collection annotations to `ResponseItem`/`BetaResponseItem`, pass through the full compacted history, and check each item’s role and concrete type.

    Tip: Store and forward unknown future item types without dropping them; remove branches that assume items are only assistant messages.

  • RetiredOpenAI Ruby SDK

    Assistants and Threads surfaces removed from the Ruby SDK

    Ruby SDK 1.0 removes `client.beta.assistants`, `client.beta.threads`, and their SDK types. OpenAI recommends switching to the Responses API with Conversations to maintain state.

    Usage details

    Who can use it: Applications upgrading to Ruby SDK 1.0 that use legacy Assistants/Threads calls.

    How to use it: Create a Conversation for each legacy thread; move instructions and tools to the Responses request, then continue subsequent calls using the `conversation` ID.

    Tip: Before migrating legacy thread history and tool results, compare the outputs of the new Responses flow using representative conversations.

  • RetiredOpenAI Ruby SDK

    Videos and Sora 2 API surfaces removed from the Ruby SDK

    Ruby SDK 1.0 removes `client.videos`, related SDK types, Sora 2 model constants, and the `/v1/videos` Batch endpoint. The migration guide states that the Videos API and Sora 2 model IDs have been retired.

    Usage details

    Who can use it: Applications upgrading to Ruby SDK 1.0 that use the Videos API or Sora 2 constants.

    How to use it: Remove `client.videos` calls, video polling/download jobs, video type annotations, and `/v1/videos` Batch inputs; switch to a supported surface.

    Tip: Search the codebase for `videos`, `Sora`, and `/v1/videos` to find dynamic usages that may not be caught by compilation.

  • RetiredOpenAI Ruby SDK

    Ruby SDK removes legacy Chat Completions model constants

    Ruby SDK 1.0 retires the `GPT_4_VISION_PREVIEW`, `GPT_3_5_TURBO_0613`, and `GPT_3_5_TURBO_16K_0613` constants. Applications need to switch to a supported model ID.

    Usage details

    Who can use it: Ruby SDK applications using these three legacy constants.

    How to use it: Replace the constants with a supported model ID; for applications using images, also verify that the new model supports image input.

    Tip: Centralize the model ID in an environment variable instead of hardcoding it; run model-specific evaluations during migration.

    ✓ Verified by official sourceModels: GPT_4_VISION_PREVIEW, GPT_3_5_TURBO_0613, GPT_3_5_TURBO_16K_0613Source: OpenAI Ruby SDK v1.0.0 sürüm notlarıSource: Ruby SDK 1.0 geçiş rehberi
  • APIOpenAI Ruby SDK

    Ruby SDK changes Live delegation ToolChoice classes

    In Live Responses delegation, function selection is now `ToolChoice::Function` and MCP selection is `ToolChoice::Mcp`. MCP selection requires `server_label`; `name` can now be omitted, sent as `nil`, or set to a specific tool name.

    Usage details

    Who can use it: Ruby SDK 1.0 applications using Live Responses delegation.

    How to use it: Replace the old selection classes with the new `Function`/`Mcp` classes; add `server_label` to the MCP configuration and handle the possibility that `name` is nil in reader code.

    Tip: In tests that serialize MCP selection, verify omission of the field and sending JSON null as separate scenarios.

  • APIOpenAI Ruby SDK

    Ruby SDK switches to a named graders API for MultiGrader

    `MultiGrader` is now initialized with a hash keyed by `graders:` and uses the grader name in `calculate_output`. Use `MultiGrader::Grader` instead of the old `MultiGrader::Graders` type.

    Usage details

    Who can use it: Applications using fine-tuning graders in the Ruby SDK.

    How to use it: Pass `graders: {ad: grader}` to `MultiGrader.new` and set `calculate_output` to the same name; update type references.

    Tip: Keep grader names stable and meaningful; compare validation results on the same examples with the old and new SDK.

  • APIOpenAI Ruby SDK

    Ruby SDK returns image sizes as strings

    `size` in image responses and generation/edit stream events is now read as a string; size constants and `Size.values` also return strings. Custom sizes remain valid.

    Usage details

    Who can use it: Applications that process Images responses or stream events in Ruby SDK 1.0.

    How to use it: Update size comparisons to use a format such as `image.size == "1024x1024"` and treat `Size.values` results as strings.

    Tip: Review serialization and validation code that assumes size values are symbols or numeric tuples.

  • RetiredAgents SDK for Python

    Migrating from the Agents SDK alpha.8 harness to the Runner flow

    Users of the previous `v0.24.0-alpha.8` harness should use `Runner.run()` or `Runner.run_streamed()` instead of `Agent.run()` or `Session.submit()`, and select `CodexProvider` in RunConfig. The `SessionState` checkpoint format is not compatible with alpha.1’s `RunState`.

    Usage details

    Who can use it: Developers migrating from the Agents SDK Python alpha.8 preview to v1.0.0-alpha.1.

    How to use it: Move imports to the `agents` package, start execution through `Runner`, finish old pending runs in the alpha.8 environment, and start a new SDK session in alpha.1.

    Tip: Do not move checkpoint files directly to the new version; keep the old and new environments running in parallel during the migration.

  • CapabilityAgents SDK for Python

    Agents SDK for Python embeds Codex Core in Python

    The preview wheel includes the unmodified Codex Core `rust-v0.166.0` component via PyO3. This removes the need to start `codex app-server` and set up JSON-RPC for every call; the same `ThreadManager` can be reused.

    Usage details

    Who can use it: Ready-to-install wheels for Agents SDK for Python v1.0.0-alpha.1; each call gets a new thread, and SDK sessions can be used for history.

    How to use it: Keep `CodexProvider` alive, share a `ThreadManager` across calls, use an SDK session for history, and call `await provider.aclose()` when finished.

    Tip: For multiple calls, share the provider instead of recreating it to reduce startup costs and resource usage.

    ✓ Verified by official sourceModels: CodexSource: Agents SDK v1.0.0-alpha.1 sürüm notları
  • AgentAgents SDK for Python

    Agents SDK for Python makes Codex an optional provider with CodexProvider

    Agents SDK v1.0.0-alpha.1 adds a preview of `CodexProvider`, which lets you run the same Python agent with Codex on a per-run basis. Codex can be selected using a saved login or by providing `api_key`; the provider can be used for the default or a named model.

    Usage details

    Who can use it: Agents SDK for Python v1.0.0-alpha.1 preview; ready-to-install packages are available only through GitHub Releases.

    How to use it: Import `CodexProvider`, create `CodexProvider()` or `CodexProvider(api_key=...)`, and pass it to the `Runner` call with `RunConfig(model_provider=provider)`.

    Tip: Keep provider selection in the run configuration to compare OpenAI and Codex providers without changing the agent definition.

    ✓ Verified by official sourceModels: CodexSource: Agents SDK v1.0.0-alpha.1 sürüm notları
  • CapabilityAgents SDK for Python

    Agents SDK supports tools and approval flows in the Codex provider

    The Codex provider supports text streaming, handoff to a specialist agent, waiting for approval before running Python tools, and typed-output validation flows. Native commands follow Codex’s sandbox and approval settings; a native approval request pauses the model call.

    Usage details

    Who can use it: Agents SDK for Python v1.0.0-alpha.1 Codex preview.

    How to use it: Run the agent with `CodexProvider`, add tools and handoffs to the SDK configuration, and use a sandbox policy that requires approval for writing or Python tools.

    Tip: Use explicit approval instead of automatic approval for tools that modify files or write to external systems.

    ✓ Verified by official sourceModels: CodexSource: Agents SDK v1.0.0-alpha.1 sürüm notları
  • AvailabilityBusinesses in ChatGPT

    Businesses in ChatGPT opens to eligible businesses outside the US

    OpenAI announced that e-commerce and online service businesses can claim their business and publish a Business Agent in ChatGPT from outside the US. This is a new geographic expansion beyond the US-only eligibility announced on November 10.

    Usage details

    Who can use it: Eligible e-commerce and online service businesses located where ChatGPT is available; no detailed country list was provided in the announcement.

    How to use it: Open ChatGPT Business Hub, claim your business, and complete verification; then add business information or a product feed, create a Business Agent, and publish it.

    Tip: Keep product feed pricing and inventory information up to date; test the agent’s responses with representative customer questions before publishing.

  • APIOpenAI Go SDK

    OpenAI Go SDK adds Live WebRTC standby support

    OpenAI Go SDK v3.105.0 adds standby support to pause inference during inactivity in Live WebRTC sessions and resume it when speech begins. The cost of idle time can be reduced while preserving the connection and session ID.

    Usage details

    Who can use it: Live WebRTC applications using OpenAI Go SDK v3.105.0; the API guide describes standby configuration for GPT-Live-1 and GPT-Live-1.1.

    How to use it: Upgrade the SDK to v3.105.0; in the WebRTC session, set `session.wake_policy.type` to `speech` and `session.idle_timeout_s` to a value between 5 and 60 seconds, and keep the microphone track open.

    Tip: Standby may be cheaper than opening a new session during long silences; however, measure the restart latency on the first brief utterance.

  • APIOpenAI Ruby SDK

    OpenAI Ruby SDK adds Live WebRTC standby support

    OpenAI Ruby SDK v0.130.0 adds support for keeping the Live WebRTC connection open and pausing inference during inactivity, then resuming it when speech arrives. This enables lower Live session costs during idle periods.

    Usage details

    Who can use it: Live WebRTC applications using OpenAI Ruby SDK v0.130.0; GPT-Live-1 and GPT-Live-1.1 sessions.

    How to use it: Upgrade the SDK to v0.130.0; use `session.wake_policy.type: speech` and `session.idle_timeout_s` between 5 and 60 seconds in the WebRTC session configuration, and keep the microphone track open.

    Tip: Monitor the `session.paused` and `session.resumed` events to make the wake-up status visible to the user.

  • CapabilityAgents SDK for Python

    Agents SDK Codex session compaction preserves history after approval pauses

    Automatic session compaction uses Codex’s local conversation history even after an approval pause. Compaction requires a separate OpenAI client or credentials.

    Usage details

    Who can use it: Agents SDK v1.0.0-alpha.1 Codex session flows; applications using compaction.

    How to use it: Continue the Codex session with an SDK session, resume it after approval, and separately configure the OpenAI client or credentials required for compaction.

    Tip: Test long tasks awaiting approval with representative history to verify that compaction does not lose context.

    ✓ Verified by official sourceModels: CodexSource: Agents SDK v1.0.0-alpha.1 sürüm notları
  • CapabilityAgents SDK for Python

    Agents SDK sandbox manifests preserve binary files when resuming

    Saved sandbox manifests now preserve inline binary files when a run is resumed. If manifests are shared, upgrading readers is recommended; `LocalFile` can be used for older SDK readers.

    Usage details

    Who can use it: Applications that save and resume Agents SDK v1.0.0-alpha.1 sandbox manifests.

    How to use it: Save and resume manifests with the new version; if older readers need to open a manifest, provide binary content with `LocalFile`.

    Tip: Before production, test reopening shared manifests with non-UTF-8 content as well.

  • AvailabilityAgents SDK for Python

    Ready-to-install wheels released for Agents SDK v1.0.0-alpha.1 on four platforms

    Ready-to-install `openai-agents` wheels for the Codex preview have been released for CPython 3.10–3.14 on Linux x86-64, Linux ARM64, macOS 11+ Apple silicon, and Windows x64. Packages are available only through GitHub Releases, not PyPI.

    Usage details

    Who can use it: For developers using CPython 3.10–3.14; requires glibc 2.28+ on Linux, and supports Apple silicon on macOS and x64 on Windows.

    How to use it: Create a virtual environment, get the wheel link for your platform, and install it with `python -m pip install --only-binary=:all: "WHEEL_URL"`.

    Tip: Verify the wheel’s SHA-256 value against the `SHA256SUMS` file in the source release and choose the wheel for your platform.

Google · Gemini

#3 Gemini 4 Neon#4 Gemini 4 Carbon#8 Gemini 4 Hydrogen#10 Gemini 4 Argon
Full history (244) →

New in this edition · 7

  • OtherGemini CLI

    Gemini CLI nightly expands home-relative paths on Windows

    Gemini CLI v0.71.0-nightly.20261121.geecccfcd5 fixes the expansion of home-relative paths containing tildes and backslashes on Windows. This allows the CLI to correctly resolve paths written relative to the user directory.

    Usage details

    Who can use it: Windows users running the `v0.71.0-nightly.20261121.geecccfcd5` nightly release.

    How to use it: Update Gemini CLI to this nightly release and use `~` and home-relative paths in Windows workspace or file paths; no separate setting is required.

    Tip: Verify the nightly release and test your automation scripts on a safe example path first.

  • OtherGemini CLI

    Gemini CLI selects the most recently active session when resuming

    The Gemini CLI nightly release now selects the most recently active session instead of the most recently started one when resuming a session. This reduces the risk of opening the wrong session in projects with parallel or intermittent work.

    Usage details

    Who can use it: Gemini CLI users running the `v0.71.0-nightly.20261121.geecccfcd5` nightly release.

    How to use it: Update to the nightly release and use `gemini --resume` or the `/resume` flow inside the CLI; the most recently active session is selected automatically.

    Tip: Send a brief status summary before closing the project so the session you want to resume remains the most recently active one.

  • AvailabilityGemini Enterprise Agent Platform

    Agent Anomaly Detection also monitors agents in external clouds

    As part of Private Preview, Agent Anomaly Detection can also monitor the behavior of agents running outside Google Cloud. When OpenTelemetry traces and logs from agents in AWS Lambda, Microsoft Foundry, and on-premises environments reach the Telemetry API in Google Cloud, anomaly analysis can be performed without accessing the runtime environment.

    Usage details

    Who can use it: Limited to Private Preview and approved/allowlisted users; telemetry from external-cloud agents must be sent to a Google Cloud project, and ADK Python 2.12.0 or later and Workload Identity Federation must be set up.

    How to use it: Request Private Preview access using the access form; authenticate the external agent with Workload Identity Federation, and configure the ADK Google Cloud exporters and OpenTelemetry variables. Enable it in the console once the agent appears in the Monitored agents list.

    Tip: Because prompt and response content can be included in telemetry, review access, data residency, and log retention policies before enabling it.

  • CapabilityGemini Enterprise Agent Platform

    Semantic governance policies review agent actions before they run

    Available as a Preview, Semantic governance policies evaluate each tool call through Agent Gateway against the user's intent and rules written in natural language; they block noncompliant calls before execution and log the decision to Cloud Logging. Anthropic, OpenAI, Mistral, and self-hosted OpenAI-compatible endpoints are also supported, in addition to Gemini.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform Preview; Agent Gateway and policy engine setup are required. Endpoints such as Gemini Developer API, Gemini platform, Anthropic Messages API, Claude, OpenAI Chat Completions, Mistral, and vLLM are supported; streaming model responses are not supported.

    How to use it: Prepare the policy engine and connect it to Agent Gateway; create a natural-language rule from the console or with the `gcloud beta ai semantic-governance-policies create POLICY_ID` command. Review Cloud Logging results in dry run first, then move to enforcement.

    Tip: Write rules and tool descriptions concretely; because LLM evaluations are probabilistic, use dry run and do not include sensitive information in denial reasons.

  • Plugin / IntegrationGemini / Google Search

    New Connected Apps integrations arrive in Gemini and Search

    Google has begun rolling out grocery ordering and payment flows through Albertsons, Safeway, Shaw’s, and Target in Gemini and Search in the US. Experian is coming to Search for credit scores and personal loan transactions, while H&R Block is available in Search for tax and refund estimates; Angi and Thumbtack integrations are expected in the coming weeks.

    Usage details

    Who can use it: Currently US only; grocery integrations are available in Gemini and Search, Experian is live in Gemini and expanding to Search, and H&R Block is available in Search; Angi and Thumbtack are still upcoming.

    How to use it: Open the Gemini app or Google Search in the US, connect the relevant Connected App account, and request your grocery, credit, or tax task in natural language; review the order and payment steps before confirming.

    Tip: Connect only the Connected App needed for each task, and verify the final total and authorizations, especially for purchases or financial transactions.

  • FeatureGemini

    GIF creation is now available in Gemini

    Gemini can now create looped GIFs of about three seconds with Gemini Omni 1.2 Flash. The official Gemini account highlights smooth motion and consistent quality across frames.

    Usage details

    Who can use it: For Google AI subscribers and free users in the US, Japan, South Korea, Australia, Canada, the United Kingdom, Germany, and France; age and account restrictions apply.

    How to use it: Enter a prompt in the main Gemini chat or the Image/Video tool; select GIF in the format selector and describe what you want. You can also upload a reference image.

    Tip: For character or product consistency, use a single reference image and a short prompt describing the start and end of the loop.

  • Pricing / PlanGoogle AI Plus / Google AI Pro

    Black Friday discounts for existing Google AI plan subscribers are now available

    Google has launched an early-access Black Friday discount of up to 50% on select Google AI plans. Google’s subscription product lead said offers for Google AI Plus and annual Google AI Pro plans are available to existing AI and storage subscribers from November 20-26, and that the exact offer varies by current plan.

    Usage details

    Who can use it: Early access is for existing AI or storage subscribers; offers for select Google AI Plus and annual Google AI Pro plans vary by country and current plan and are valid through November 26.

    How to use it: Open the Google AI plans page with your Google account, select the Black Friday offer shown, and verify the discount and promotion end date on the checkout screen.

    Tip: Standard pricing may apply after the promotion ends, so check the renewal date and cancellation terms in the Google One subscription screen in advance.

Meta · Meta AI

Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 8

  • APIMeta Wearables Device Access Toolkit for iOS

    DAT iOS stream errors and photo timeout are more informative

    If the glasses can't start a stream due to heat, low battery or the peak power limit, `Stream` now returns `thermalHot`, `batteryLow` or `peakPowerLimit`, respectively. `Stream.capturePhoto(format:)` now reports a failed photo within 10 seconds instead of 30.

    Usage details

    Who can use it: iOS apps using camera streaming or photo capture with DAT iOS 1.1.0.

    How to use it: Handle the new `StreamError` states separately in your stream error handler; for photo capture, implement a retry or user message based on the 10-second timeout.

    Tip: For `thermalHot` and `peakPowerLimit`, cool the device or address its power state instead of retrying immediately; this prevents unnecessary error loops.

  • APIMeta Wearables Device Access Toolkit for iOS

    New error and cancellation handling added to DAT iOS registration and session flows

    The SDK reports session creation before registration with `DeviceSessionError.notRegistered`, and unauthorized registration attempts with `RegistrationError.unauthorized`. With `cancelRegistration(returnToMetaAI:)`, the app can choose whether to return to Meta AI after cancellation; when the app is deregistered, active sessions also transition to `stopped`.

    Usage details

    Who can use it: Meta Wearables integrations using DAT iOS 1.1.0; developers using registration and app authorization flows.

    How to use it: Before creating a session, wait for `registrationState` to be `.registered`; handle the `notRegistered` and `unauthorized` states. On user cancellation, call `cancelRegistration()`; for setup errors, call `cancelRegistration(returnToMetaAI: false)` with an explicit user choice.

    Tip: During deregistration or re-registration, wait for the `stopped` event and create a new `DeviceSession` instead of reusing open sessions.

  • APIMeta Wearables Device Access Toolkit for iOS

    DAT iOS MockDevice testing APIs updated

    Version 1.1.0 adds services that simulate fake glasses requiring a DAT app or firmware update, the `MockDeviceTestClient.kit` flow for controlling `MockDeviceKit` from the UI test process, and `MockDeviceTestGlasses.services` access. If the test server cannot be started, a specific `MockDeviceKitError.testServerUnavailable` error is now returned.

    Usage details

    Who can use it: iOS teams using the DAT iOS developer preview for integration and UI testing without physical Meta AI glasses.

    How to use it: Use `MockDeviceTestClient.kit` with `kit.pairGlasses(model:)`; access device services through `MockDeviceTestGlasses.services`. Enable update prompts with `setDATAppUpdateRequirementEnabled(_:)` and `setFirmwareUpdateRequirementEnabled(_:)`, then disable them when testing is complete.

    Tip: Set the firmware requirement before turning on the glasses, and the DAT app requirement before the next session starts; this makes tests deterministic.

  • APIMeta Wearables Device Access Toolkit for iOS

    Pause and resume session APIs added to the DAT iOS SDK

    Version 1.1.0 adds the `DeviceSession.pause()` and `DeviceSession.resume()` methods. An active session can transition from `started` to `paused` and then back to `started`, with transitions tracked through the session state.

    Usage details

    Who can use it: DAT iOS 1.1.0 users in the developer preview building iOS apps for Meta AI glasses.

    How to use it: Start a `DeviceSession`; call `pause()` while it is in the `started` state and `resume()` while it is in the `paused` state, then observe the `started → paused → started` transitions through `stateStream()`.

    Tip: Enable or disable pause/resume buttons in the UI based on the current session state, rather than assuming calls in invalid states are no-ops.

  • Plugin / IntegrationMeta Wearables Device Access Toolkit for iOS

    MCP server added to DAT iOS for controlling simulated devices with AI coding tools

    The SDK adds a bundled MockDevice MCP server for validating simulated device behavior without physical glasses. The server runs on the Python 3.8+ standard library and registers automatically when the plugin is installed.

    Usage details

    Who can use it: Developers using DAT iOS 1.1.0 to test with MCP-compatible AI coding tools such as Muse Code, Claude Code, Codex or similar.

    How to use it: Install the relevant plugin from the `meta-wearables-dat-ios` repository; wait for the MCP-compatible coding tool to register the server automatically, then run simulated glasses flows through MockDevice.

    Tip: First validate pairing, permission and update prompts without physical hardware to reduce the number of device lab tests.

  • OtherATLAS

    Meta FAIR announces ATLAS, an open-source Lean mathematics library

    Meta's FAIR team announced ATLAS as an open-source project at the Open Source AI For Science Summit. ATLAS draws on 26 open textbooks and more than 45,000 computer-verified definitions and results for human- and machine-assisted theorem proving.

    Usage details

    Who can use it: Researchers and developers; the GitHub repository and project based on Lean 4.35.0 are publicly available.

    How to use it: Clone the repository: `git clone https://github.com/facebookresearch/atlas-lean.git`. To build the archived first release, run `cd atlas-lean/v1 && lake build`; for checks on the current root project, run `scripts/check.sh`.

    Tip: Check the differing status and licenses of v1 and v2, and base new development on the repository's current v2 structure.

  • RetiredMeta Wearables Device Access Toolkit for iOS

    Legacy flat APIs deprecated in DAT iOS MockDevice tests

    `MockVoiceInvocationKit.sendLaunchAppAction()`, `sendIncompleteAction()` and `hasConnectedClients` are now deprecated; they are replaced by `simulateLaunchAppAction()`, `simulateIncompleteAction()` and `hasConnectedApps`. The flat device methods in `MockDeviceTestClient` are also deprecated in favor of the `kit`-based flow.

    Usage details

    Who can use it: Developers using MockDevice or UI tests with DAT iOS 1.1.0.

    How to use it: Use the new `simulate...` methods for voice action simulations; in the test client, make calls through the glasses and service objects after `kit.pairGlasses(model:)`.

    Tip: Don't defer compiler warnings; moving from the old flat methods to the `kit` flow also makes new service and permission simulations accessible.

    ✓ Verified by official sourceSource: Meta Wearables DAT iOS CHANGELOG.md
  • RetiredMeta Wearables Device Access Toolkit for iOS

    CrashReporting OptOut setting removed in DAT iOS

    Version 1.1.0 removes the `MWDAT > CrashReporting > OptOut` setting. Use `MWDAT > Analytics > OptOut` to control analytics and crash reporting together.

    Usage details

    Who can use it: iOS apps moving to DAT iOS 1.1.0 that use the old CrashReporting setting in Info.plist.

    How to use it: Remove the `MWDAT/CrashReporting/OptOut` key from Info.plist; set `MWDAT/Analytics/OptOut` to `true` if you want to disable both data collection behaviors.

    Tip: Recheck Info.plist and your privacy documentation after upgrading the SDK; don't assume the old key is still effective.

    ✓ Verified by official sourceSource: Meta Wearables DAT iOS CHANGELOG.md

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 11

  • CapabilityQwen Code

    Qwen Code adjusts context compression to fit the model window

    Qwen Code compression now reduces or rejects summary requests that would not fit in the model window, and attempts recovery once after a context-length rejection. This reduces the chance of compression failing and entering a loop during long sessions.

    Usage details

    Who can use it: CLI, ACP, and Web Shell sessions that use context compression in Qwen Code v0.28.3.

    How to use it: Run the normal context compression flow in a long session; if the request does not fit in the window, let Qwen Code reduce it or make its one-time retry.

    Tip: In sessions approaching the compression limit, clear unnecessary tool output and keep important decisions in concise summaries.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • CapabilityQwen Code Web Shell

    Qwen Code prevents sessions with running background jobs from being deleted

    Web Shell now prevents any session with a running background job from being deleted and displays the reason. The session becomes available for deletion again once the job finishes.

    Usage details

    Who can use it: Qwen Code v0.28.3 Web Shell sessions.

    How to use it: Start the session deletion action; if a background job is running, stop the job or wait for it to finish, then select delete again.

    Tip: Export outputs and artifacts before deleting long-running workflows.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • Plugin / IntegrationQwen Code

    Qwen Code adds an experimental WeCom Managed Channel adapter

    Qwen Code v0.28.3 adds an experimental channel adapter that routes private WeCom chats to connected Managed Agent Sessions. Replies are sent back to WeCom over an authenticated WebSocket.

    Usage details

    Who can use it: Users running Qwen Code v0.28.3 who connect WeCom with Managed Agent; experimental feature.

    How to use it: Enable the experimental WeCom adapter in Qwen Code Managed Channels settings, configure the authenticated WebSocket connection, and link the private chat to a connected Managed Agent Session.

    Tip: Test connection drops and reply routing in WeCom private chats before using the experimental adapter for production traffic.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • AvailabilityTogether AI Serverless Inference

    Qwen4-Max launches on Together AI Serverless

    Together AI has made Alibaba's Qwen4-Max model available on Serverless Inference with the model ID `Qwen/Qwen4-Max`. The discounted prices announced through November 22, 2026 are $1.05 per million input tokens, $0.0525 per million cached input tokens, and $3.15 per million output tokens; Together's announcement also states that requests must use streaming and that thinking is enabled by default.

    Usage details

    Who can use it: Together AI Serverless Inference; users with an API key. The model is available only as serverless, and the announced discount is valid through the end of 2026-11-22.

    How to use it: Create a Together API key and send a request to `https://api.together.ai/v1/chat/completions` with `model: Qwen/Qwen4-Max` and `stream: true`; in the Python SDK, use `client.chat.completions.create(..., stream=True)`. If needed, turn off thinking with `reasoning: {enabled: false}`.

    Tip: Take advantage of automatic prompt caching for large repository prompts; however, check that the repository contains no secret keys before sending it to Together.

  • CapabilityQwen Code

    Qwen Code caps combined async Hook context at 10,000 characters

    The combined async Hook context block is capped at 10,000 characters before it is sent to the model. This helps prevent Hook output from growing excessively and causing the model request to be rejected.

    Usage details

    Who can use it: CLI, ACP, and Managed Agent flows that use async Hook in Qwen Code v0.28.3.

    How to use it: Keep the necessary information under 10,000 characters when generating async Hook output; Qwen Code v0.28.3 applies the limit automatically.

    Tip: Structure Hook output with concise status, error, and next-action fields instead of raw logs.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • AvailabilityQwen Code Desktop ve CLI

    Qwen Code completes Russian interface coverage for Desktop and CLI

    Qwen Code v0.28.3 completes Russian interface coverage in Desktop and CLI and keeps long account menu labels on a single line. This localization update fills in previously missing product and settings text for Russian-speaking users.

    Usage details

    Who can use it: Users of the Russian interface in Qwen Code Desktop and CLI.

    How to use it: Select Russian/ru in the Qwen Code Desktop or CLI language settings and upgrade to v0.28.3.

    Tip: Check the newly translated experimental and visionBridgeTimeoutMs text so you can use the exact setting name in support requests.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • Plugin / IntegrationQwen Code

    Qwen Code ties DingTalk pending messages to the registration ID

    The DingTalk Managed Channel persistent ledger is now tied to the robot application's registration ID. This prevents pending messages from being accidentally replayed under a different registration for the same application.

    Usage details

    Who can use it: DingTalk Managed Channel installations using Qwen Code v0.28.3.

    How to use it: Upgrade Qwen Code to v0.28.3; when you reconnect the DingTalk channel with the same registration ID, the pending message queue is preserved against that ID.

    Tip: When re-registering the robot application, don't mix up the old and new registration IDs; otherwise, pending messages may not reach the session you expect.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • CapabilityQwen Code

    Qwen Code can write valid UTF-8 BOM-prefixed text with LocalDirectory

    LocalDirectory write operations now accept text that starts with a valid UTF-8 BOM. The behavior of rejecting invalid unpaired surrogate characters remains unchanged.

    Usage details

    Who can use it: Managed Agent and CLI flows using LocalDirectory in Qwen Code v0.28.3.

    How to use it: Use the LocalDirectory write tool to send valid UTF-8 text containing a BOM; do not send invalid unpaired surrogate content.

    Tip: For BOM-prefixed files generated by Windows, verify the file encoding separately after writing.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • SettingQwen Code Web Shell

    Qwen Code groups approval scopes in a dropdown

    Approval scope options are now grouped in a dropdown in the Web Shell approval interface. This makes the interface more compact and easier to navigate when approvals have multiple scopes.

    Usage details

    Who can use it: Qwen Code v0.28.3 Web Shell approval screen.

    How to use it: When a tool or action approval appears, open the approval scopes dropdown in Web Shell and select the required scope.

    Tip: Before granting persistent permission, choose the option with the narrower scope; review file and shell permissions separately in particular.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • CapabilityQwen Code

    Qwen Code correctly matches Windows drive and UNC permission paths

    Permission rules now correctly match absolute paths with Windows drive letters and absolute UNC paths. This includes Always Allow entries, reducing unnecessary repeated confirmations for the same file.

    Usage details

    Who can use it: Windows users running Qwen Code v0.28.3.

    How to use it: When adding a permission rule, save an absolute Windows drive path or UNC path under Always Allow; Qwen Code will use that path for subsequent matches.

    Tip: For shared folders, use the UNC path instead of a drive letter to make the rule behave more predictably across different machines.

    ✓ Verified by official sourceSource: Qwen Code v0.28.3 sürüm notları
  • ModelQwen Hugging Face model kartları

    Qwen Hugging Face cards point five older models to newer versions

    On November 20, the Qwen team added a `new_version` field to the metadata of five older model cards: Qwen2-VL-2B-Instruct → Qwen3-VL-2B-Instruct, Qwen3.6-27B → Qwen3.8-27B, Qwen3.5-27B → Qwen3.6-27B, Qwen3.5-35B-A3B → Qwen3.6-35B-A3B, and Qwen2-1.5B-Instruct → Qwen2.5-1.5B-Instruct. Hugging Face now shows a new-version notice to people who open an older card. No changes to the weights were announced.

    Usage details

    Who can use it: Anyone opening the relevant Qwen model cards on Hugging Face; open-weight models.

    How to use it: Open the older model card, follow the new-version notice at the top to the target card, and update the model ID to the new name in your Transformers, vLLM, or SGLang configuration.

    Tip: Before changing the model ID, compare the new version's memory requirements, context length, and license terms using your own evaluation set.

    ✓ Verified by official sourceModels: Qwen2-VL-2B-Instruct, Qwen3-VL-2B-Instruct, Qwen3.6-27B, Qwen3.8-27B, Qwen3.5-27B, Qwen3.5-35B-A3B, Qwen3.6-35B-A3B, Qwen2-1.5B-Instruct, Qwen2.5-1.5B-InstructSource: Qwen kuruluş etkinlik akışı (Hugging Face)Source: Qwen2-VL-2B-Instruct new_version tartışmasıSource: Qwen3.6-27B new_version tartışması

Watchlist

Muse Wonder Full

50
Meta · Meta AI

Tied with Gemini 4 Argon at 50 points, but just outside the TOP 10 because it ranks 11th in the Artificial Analysis table. Meta’s strongest Muse model; no public API (no API provider is shown on Artificial Analysis), and available to Muse Code users.

Kimi K3.1 Pro

46
Moonshot AI · Kimi

In 12th place; still the strongest open-weight model. API $2 / $10, with weights on Hugging Face, Amazon Bedrock and, according to Artificial Analysis, 16 API providers. Kimi Code 2.7.0 introduced the `kimi server` command, which works without opening a browser; Desktop builds 1.4.0 and 1.5.0 Canary were released on GitHub without release notes.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; ranked 13th with 44 points. In preview on Google Agent Platform as `muse-wonder-1.0`.

Qwen4 Max

43
Alibaba · Qwen

Released on November 16; ranked 14th with 43 points. `qwen4-max` on Model Studio (1M context, Singapore pricing $1,50 / $4,50), QwenCloud Token Plan, QwenWork, QwenStudio and OpenRouter (`qwen/qwen4-max`). According to Together AI’s November 21 announcement, it is also available on Serverless Inference as `Qwen/Qwen4-Max` (not separately confirmed by Alibaba). Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

Anthropic’s fast, low-cost model at $0.10 / $0.50. According to Meta, the new Muse Ember delivers the same score at roughly one-third the task cost.

Muse Ember

36
Meta · Meta AI

Released on November 9. Meta’s smallest Muse model; $0,10 / $0,50, 1M context, `muse-ember-1.0` on Meta Model API and available on all Muse Code plans.

Sources