YZ·RADARyz.mikail.net
Daily edition · 21 November 2026

TOP 10 unchanged; Tencent Hy4 becomes open-weight, Claude computer use comes to Team

No changes in the TOP 10 today either; the rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are the same as 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 ties Argon at 50 points but remains outside the list, in 11th place. The watchlist is unchanged too: 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). Among the new models added by Artificial Analysis, Thinking Machines’ Inkling 2 Mini (November 20) scored 20 points; it is an Apache 2.0-licensed MoE with 26B total and 5.4B active parameters that can run on small devices, but it is not strong enough to make the list. Tencent Hy4, whose weights were released on November 20, is also absent from the top 16 in the Artificial Analysis table. LMArena’s live page was again blocked by CAPTCHA; the latest accessible snapshot is dated November 11, with Gemini 4 Carbon, Gemini 4 Argon, and Muse Wonder in the top three.

Anthropic’s news of the day is that the computer use beta is now available on Team and Enterprise plans. Claude can view and operate permitted desktop apps; the rollout to organizations is gradual. It’s available in the new Claude Desktop experience and on web and mobile when Computer access is enabled. The other 40 entries are from Claude Code v2.1.330, released on November 20. Claude Code can now sign in to organizations managed through the Anthropic Enterprise Admin Console and apply the organization’s cloud provider or gateway settings. The built-in Easel mode automatically creates charts when requested and below MCP metric results; it can be disabled in settings. Scheduled tasks and `/loop` in a running session can now run for up to 30 days instead of seven; tasks created before this release still have a seven-day limit. The automatic `max_tokens` retry removed in 2.1.329 has returned. More `rm` commands targeting the root, home, or working directory now require confirmation. VS Code adds `/fork`, `/workflows`, and the `claudeCode.escapeStopsClaude`, `claudeCode.pinPrompt`, and `claudeCode.contextUsageThreshold` settings. Self-hosted runners now support a repository-local startup directory, two Prometheus metrics, and up to four concurrent `spawn-runner` hooks. At OpenAI, ChatGPT Space has started rolling out globally to Plus users aged 18 and over; users can work with ChatGPT and teammates on pages, presentations, and sites. The redesigned mobile app with Home and Explore tabs is first rolling out to Pro users and a small group of Plus users; the European Economic Area, the United Kingdom, Switzerland, and Korea are excluded. Enterprise workspace owners can set person-specific SSO exceptions for external members. In the API, `gpt-live-1.1` launched for multilingual voice conversations with more accurate transcription and more consistent pronunciation; WebRTC sessions can switch to standby when inactive. Active time costs $0,05 per minute; standby costs $0,02 per minute. Node SDK 7.61 and 7.62 added support for this.

Google’s 10 entries are mostly enterprise-focused. In Gemini Enterprise, the Microsoft 365 data store is generally available, bringing together SharePoint, OneDrive, Excel, Outlook, calendars, contacts, tasks, Teams, and Microsoft Entra data in one source. Search, read, create, and update actions are available when enabled by an administrator. Third-party data stores such as AirOps, Airtable, Asana, GitLab, Linear, Stripe, Zapier, and Zoho Projects have also exited preview. Supervised fine-tuning, continuous tuning, and tuning checkpoints for Gemini 3.8 Flash are generally available. Workspace Studio support for US and European Data Regions is generally available; with the new Collaboration steps setting, administrators can limit flows’ ability to share information with others. The setting is on by default. Google Vids stores the images, videos, and audio used in videos generated with Gemini Omni in the project’s Assets section. A collapsible side panel is rolling out globally on the desktop Google Search homepage, making it easier to return to recent AI searches. Gemini CLI 0.70.0 is now stable, and 0.71.0-preview fixed a PowerShell freeze. Tencent released Hy4’s weights: a multimodal model with 770B total and 49B active parameters and a 1M context window. The instruction-tuned Hy4, FP8-quantized Hy4-FP8, and pretrained Hy4 base variants are available on Hugging Face, ModelScope, and GitCode. Official serving recipes for vLLM and SGLang have been published. Hy4 is available on Together AI as `tencent/Hy4` at $0,83 / $2,50; Hy4 flash is free on OpenCode for 10 days starting November 18. LoopForge added platform-wide Git SSH host trust records and a token and cost panel for scheduled runs. Moonshot’s 10 entries cover Kimi Code fixes and Hugging Face cards: the `kimi server` command in Kimi Code 2.7.0 starts a local server without opening a browser; 2.7.1, 2.7.2, and Desktop 1.4.1/1.4.2 fixed the selected thinking effort value resetting and a fork error after steering. The Kimi K3 card on Hugging Face now points to K3.1 Pro, and the Kimi K2.5 card points to K2.6.

Of Alibaba’s 64 entries, 28 come from Qwen Code v0.28.3, released on November 21: a Microsoft Teams channel extension, an optional screen-reader view for the CLI, headless background workflow execution, waking a linked session when a tracked PR has a CI failure or conflict, a workspace artifact library, skipping a model whose capacity is full during the same turn, and sending local media through a QQ bot. On the Qoder side, Subagents now run in the background by default in the CLI, and a `protocolVersion` option has been added for MCP servers. Qoder IDE adds in-chat search with Ctrl/Cmd+F and a policy that restricts sign-in on managed devices to specified organizations. QoderWake is now connected to WeCom channel posts, and `mindlab/macaron-v1.1-venti-pro` was added to Model Studio Singapore. At Meta, UMA Nitro accelerates atomic energy and force calculations on Ampere, Hopper, and Blackwell GPUs; UMA Playground is now available to high school chemistry teachers. Starting in February 2027, WhatsApp marketing messages that cannot be delivered because of template quality or max price will return code 131066 instead of 131049. According to the Xiaomi Token Plan FAQ, packages cover six models; team packages can be purchased as one-time subscriptions for 1, 3, 6, or 12 months, and when the quota runs out, only that seat stops. xAI removed `grok-build-0.1` from the API on November 19; the official note did not name a successor model. DeepSeek’s chat-demo v0.1.3 resizes images with a longest edge over 4,096 pixels before upload. Z.ai had no new changes between November 19 and 21. A total of 155 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 no change in score or ranking. Available on all paid Claude plans; API model `claude-fable-5-2`, $8 / $40, with thinking always adaptive. The Claude Meetings beta is available on Team and Enterprise plans; since November 20, the computer use beta has also been gradually rolling out to these plans.Released: 9 November 2026

  2. 02

    GPT-6.2 Astra

    OpenAI · ChatGPT=
    62pts

    Released on November 17; remains in second place with 62 points. Costs $4,00 per task at the Max setting. API model `gpt-6.2-astra`, $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 hard 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. Costs $6,44 per task at the xhigh setting. API model `gemini-4-neon` (paid tier only), $3 / $15, 1M context; has low, medium, high, and xhigh thinking levels. Generally available on Gemini Enterprise Agent Platform; according to Artificial Analysis, also available to Google AI Pro and Ultra subscribers in the Gemini app and Antigravity. It can also be selected in Gemini Enterprise Workflow Builder and BigQuery generative AI features.Released: 18 November 2026

  4. 04

    Gemini 4 Carbon

    Google · Gemini=
    60pts

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

  5. 05

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Score (59) and ranking are unchanged. Anthropic’s flagship Opus model; available on Claude Pro/Max/Team/Enterprise and through the API, as well as via 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. Costs $1,01 per task at the Max setting, one-fourth of Astra’s cost. API model `gpt-6.2-sol`, $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 on Claude.ai’s free plan as well; 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. Costs $3,16 per task at the xhigh setting. API model `gemini-4-hydrogen` is also available on the free 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 images and computer use. 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 with Google AI Pro/Ultra and on the paid API; promotional price of $2 / $10 through November 30. Tied with Muse Wonder Full at 50 points; kept tenth place because it ranks ahead of Muse Wonder Full in the Artificial Analysis table.Released: 30 September 2026

Key signals

  • The TOP 10 is unchanged: Claude Fable 5.2 (65) leads, with all scores and rankings the same as yesterday.
  • Tencent Hy4 (770B/49B, 1M context) is now open-weight; available on Together AI for $0,83 / $2,50.
  • Claude’s computer use beta is available on Team and Enterprise; `/loop` in Claude Code 2.1.330 now runs for up to 30 days.
  • Microsoft 365 and third-party data stores are GA in Gemini Enterprise; Gemini 3.8 Flash fine-tuning is GA.
  • ChatGPT Space is rolling out to Plus; the API adds GPT-Live-1.1 and WebRTC standby at $0,02 per minute.
  • 155 new developments: Alibaba 64, Anthropic 41, Tencent 14, Google 10, Moonshot 10, OpenAI 6, Xiaomi 4, Meta 3, DeepSeek 2, xAI 1.

Worth trying today

01

Set up longer loops with Claude Code 2.1.330

Run `claude update` to upgrade to 2.1.330. Scheduled tasks and `/loop` in a running session can now run for up to 30 days, but tasks you created before this release remain subject to the seven-day limit; stop and recreate monitoring loops in `/tasks` if you want them to run longer. If you use a gateway with a lower output limit, automatic retries for `max_tokens` overages are back, so the `CLAUDE_CODE_MAX_OUTPUT_TOKENS` setting recommended yesterday is no longer required. In VS Code, use `/fork` to branch a conversation from the last response and `/workflows` to view running workflows; to prevent Escape from stopping Claude, turn off `claudeCode.escapeStopsClaude`. If you’re on Team or Enterprise, try the computer use beta with a single low-risk desktop app first when it becomes available to your organization.

Anthropic · Claude
02

Connect your Microsoft 365 data to Gemini Enterprise

If your company uses Microsoft 365, a Gemini Enterprise administrator can now connect SharePoint, OneDrive, Excel, Outlook, calendars, contacts, tasks, Teams, and Entra data from one source using the generally available Microsoft 365 data store. Start by enabling search and read actions only; try create and update actions with a small pilot group. GCC High and DoD tenants need to use their own Microsoft Entra application. If you’ve tried the Asana, GitLab, Linear, Stripe, or Zapier connectors in preview, you can now move them into production because they are generally available. If you use Workspace Studio, review the Collaboration steps setting in the Admin console; it’s on by default. Turn it off for departments where flows should not share files.

Google · Gemini
03

Lower idle costs for voice agents with GPT-Live standby

For voice assistants where users may be silent for long periods (call centers, kiosks, voice guides), use WebRTC standby instead of closing and reopening the session: the session stays open, inference pauses when inactive, and conversation resumes in the same context. Active time costs $0,05 per minute, compared with $0,02 per standby minute. Set the model to `gpt-live-1.1` for multilingual conversations; this model works only on `v1/live/sessions`, not Chat Completions or Responses. If you use Node.js, upgrade the SDK to 7.62.0 or later and apply the standby configuration in OpenAI’s Live guide.

OpenAI · ChatGPT
04

Try Tencent Hy4 on a hosted endpoint first

Hy4 is now open-weight, but with 770B total parameters, running it on your own servers requires a multi-GPU cluster. First, test your own tasks with the `tencent/Hy4` model on Together AI ($0,83 input / $2,50 output, 1M context, tool calling, and JSON output). If the results are satisfactory, serve the Hy4-FP8 variant to reduce memory use, following the official vLLM or SGLang recipes. Hy4 base is not an instruction-tuned chat model; choose instruction-tuned Hy4 for chat and agent tasks. For lighter workloads, Hy4 flash is free on OpenCode for 10 days starting November 18.

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 · 41

  • AvailabilityClaude

    Computer use beta is now available on Team and Enterprise plans

    Computer use is now available in beta on Team and Enterprise plans. Claude can view and interact with permitted desktop apps; rollout to organizations is gradual.

    Usage details

    Who can use it: Team and Enterprise organizations; with the new experience in Claude Desktop, and when Computer access is enabled on web and mobile.

    How to use it: A Team or Enterprise administrator should complete the Computer use setup in the help center; users can enable Computer access in Claude Desktop or on web/mobile.

    Tip: For the first try, grant access only to the necessary desktop apps and keep the scope narrow.

    ✓ Verified by official sourceSource: Claude Help Center sürüm notları
  • FeatureClaude Code

    Easel added as a built-in mode that automatically draws charts

    The built-in mode called Easel draws charts on request and below MCP metrics results. Automatic chart generation can be turned off in Claude Code settings.

    Usage details

    Who can use it: Pro and Max on macOS and Linux; Team and Enterprise where available.

    How to use it: Request a chart or get an MCP metrics result; to turn off automatic charts, change the Easel setting in `/config`.

    Tip: For dense metrics output, ask for the chart first and then the raw data for quick verification.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AvailabilityClaude Code

    Direct sign-in added for organizations on Enterprise Admin Console

    Claude Code can now sign in to organizations managed by Anthropic Enterprise Admin Console. After sign-in, it applies the organization’s settings and configured cloud provider or gateway.

    Usage details

    Who can use it: Organizations managed by Anthropic Enterprise Admin Console.

    How to use it: Run `claude auth login` in the terminal and complete organization sign-in; then restart the session and verify that the organization settings have loaded.

    Tip: When troubleshooting, check the provider or gateway selected by the organization before local settings.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • SettingClaude Code

    Approval requirements expanded for rm commands targeting root and working directories

    More `rm` commands targeting the root, home, a top-level directory, or the working directory now require approval; in environments where a response cannot be provided, the command is rejected.

    Usage details

    Who can use it: Interactive and non-interactive Claude Code sessions.

    How to use it: When the command requests approval, review the target path and deletion scope, then explicitly approve it; account for the possibility that the command may be rejected in headless workflows.

    Tip: Limit deletions to a narrow target path within the project and an explicit file list.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AgentClaude Code

    Subagent model fallback and Bedrock region now propagate to subflows

    Subagents without a specified model now inherit the fallback model in use; forks and subagents they launch also use the relevant Bedrock region.

    Usage details

    Who can use it: Claude Code sessions that use subagents and forks; region information matters on Bedrock.

    How to use it: Set the session model and fallback setting, then launch a subagent; verify that the subflow runs with the expected model and region.

    Tip: Account for subagents switching to the fallback model when estimating cost and latency.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AgentClaude Code

    Background agents and Remote Control stopping status are more consistent

    Fixed stopped background subagents restarting and Claude continuing to wait for them when they were nearly finished. Also fixed Diff and pull request buttons disappearing in Remote Control mobile apps and stopped sessions appearing to be running.

    Usage details

    Who can use it: Claude Code background agents and desktop/mobile sessions using Remote Control.

    How to use it: Stop a background agent and monitor its status; in a conversation opened mid-session, check the Diff or pull request view in the Remote Control mobile app.

    Tip: In automations, check the agent status event as well as the running indicator.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Scheduled tasks and /loop in active sessions can now run for up to 30 days

    Recurring scheduled tasks and `/loop` can now run for up to 30 days instead of seven, as long as the session remains active. Tasks created before this release retain the old seven-day duration.

    Usage details

    Who can use it: Active sessions in Claude Code and newly created tasks.

    How to use it: Create a new recurring task in a session or use `/loop`; watch for Claude Code’s notification when the duration ends.

    Tip: For long-running tasks, break the work into small, observable steps rather than relying on the 30-day window.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APIClaude Code

    Disk, model, and fast mode diagnostics fixed in claude -p and SDK flows

    When redirecting `claude -p` output to a file on a full disk, it now prints an error and exits with code 1. Also fixed `fastMode` ignoring pricing when the model comes from an environment variable or settings file on Google Cloud Agent Platform.

    Usage details

    Who can use it: Flows using headless `claude -p` and Google Cloud Agent Platform.

    How to use it: Check the exit code in scripts that redirect output to a file; verify `--settings '{"fastMode":true}'` on Agent Platform.

    Tip: Store stdout and stderr separately in CI, and don’t interpret a disk-full error as success.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationClaude Code

    Claude in Chrome reconnects, and Bedrock sessions retain their pinned model

    Fixed Claude in Chrome incorrectly remaining not connected after a brief connection issue. Amazon Bedrock terminal sessions also no longer ignore the pinned Opus model when no model is selected.

    Usage details

    Who can use it: Claude in Chrome and Claude Code sessions using Amazon Bedrock/Mantle.

    How to use it: With Chrome open and signed in, restart the `/chrome` flow from Claude Code; in Bedrock, keep the pinned model setting when the default model is empty.

    Tip: Don’t overlook the provider fallback notice; check which model the call actually uses.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Chart, MCP termination, and context visibility fixed in Code Mode

    Fixed Code Mode’s long pauses in projects with hundreds of files, failures in the built-in chart tool, and MCP tool requests to end a turn from a script being ignored. `/context` now counts system tools more accurately before the first response.

    Usage details

    Who can use it: Claude Code Code Mode, especially auto and bypass permissions flows.

    How to use it: Enable Code Mode and call a chart or MCP tool; use `/context` to check the session tool count.

    Tip: Set path-scoped rules and Read deny restrictions in advance when scanning very large files.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Fixed concurrent settings saves and resume flows with invalid timestamps

    Fixed an issue where user settings could be lost when multiple Claude Code processes saved settings, and an issue where `--continue` or `--resume` could open an empty session or fail if the last message had an invalid timestamp.

    Usage details

    Who can use it: Users running multiple Claude Code processes or resuming older sessions.

    How to use it: Upgrade Claude Code to 2.1.330; for affected sessions, select the session from the `--resume` list and verify that the transcript reopens.

    Tip: Reduce the number of parallel processes writing to the same settings file during updates.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationClaude Code

    Concurrent plugin tasks and Claude settings backups are preserved

    Fixed issues where plugins disappeared from the list when multiple plugin commands ran at once, dependency and metadata loss during concurrent task updates, and a bad `.claude.json` backup on a full disk overwriting a good backup.

    Usage details

    Who can use it: Claude Code sessions using plugins and tasks.

    How to use it: Update to 2.1.330; check `/plugin` and `/tasks` after running parallel plugin commands and task updates.

    Tip: Monitor low-disk-space warnings and leave room for backups in the Claude configuration directory.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AvailabilityClaude Code

    Sign-in, connector removal, and account hold errors are clearer

    Fixed the unnecessary rejection of `forceLoginOrgUUID` initialization after a temporary connection error, ambiguity around removing a claude.ai connector, and generic errors shown for accounts on hold.

    Usage details

    Who can use it: Enterprise sign-in enforcement, claude.ai connectors, and account hold flows.

    How to use it: Try again when prompted to sign in; disable the connector from `/mcp` or remove it on claude.ai; if the account is on hold, follow the displayed appeal link.

    Tip: Include the reason and appeal link now shown, rather than a generic error, in support tickets.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Faster shell workflows on Linux and after Git Bash startup on Windows

    Fixed Bash delays and lost diffs on Linux in repositories with large ignored directories, as well as command slowdowns on Windows that persisted after Git Bash’s slow initial startup.

    Usage details

    Who can use it: Claude Code users on Linux, Windows, and WSL.

    How to use it: Update Claude Code to 2.1.330; rerun Bash commands in repositories containing large ignored directories such as `node_modules`.

    Tip: Narrowing repository ignore rules reduces both search and diff costs.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationClaude Code

    LSP, /mcp, and plugin servers are classified more accurately

    An issue causing concurrent calls and diagnostics to fail while a slow LSP was starting has been fixed. `/mcp` now lists plugin servers under Plugin MCPs and other servers separately, and identifies the plugin name on the server page.

    Usage details

    Who can use it: Claude Code sessions using LSP and connecting plugin or other MCP servers.

    How to use it: Open `/mcp` and check that servers appear under Plugin MCPs and Other MCPs; try multiple file operations while an LSP server is starting.

    Tip: If a server has an issue, first read the reason on its server page, then use reconnect.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APIClaude Code

    Automatic retries restored for max_tokens limit errors

    Claude Code 2.1.330 restored the automatic retry removed in 2.1.329 for requests that exceed a model or gateway’s `max_tokens` limit.

    Usage details

    Who can use it: Workflows using Claude Code through a gateway, custom base URL, or provider.

    How to use it: Update Claude Code to 2.1.330; if the gateway uses a lower limit, set `CLAUDE_CODE_MAX_OUTPUT_TOKENS` to that limit.

    Tip: Explicitly matching the gateway limit with a fixed setting reduces retry-related latency.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationClaude Code

    MCP resources, connector names, and server lifecycle are more reliable

    MCP resources from claude.ai connectors and plugin servers are now added when referenced with `@` in the terminal prompt. Issues that caused different connector names sharing the same URL to disappear, or the shared MCP server to shut down when a subagent finished, have been fixed.

    Usage details

    Who can use it: Sessions using claude.ai connectors, plugin MCP servers, and subagents.

    How to use it: Reference the MCP resource with `@`; if multiple connectors use the same URL, check their names in the `/mcp` list.

    Tip: Use unique connector names for the same server URL.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APIClaude Code

    commandPrinted and commandFailed added to plugin installation JSON output

    `claude plugin install --json` now reports whether the installation command wrote output to stderr and whether it ended with an error, was terminated, or timed out.

    Usage details

    Who can use it: Scripts that automate plugin installation using JSON output.

    How to use it: Read the `commandPrinted` and `commandFailed` fields in the output of `claude plugin install --json` and evaluate the CI result accordingly.

    Tip: Do not treat `commandPrinted` alone as a failure; check the actual status alongside `commandFailed`.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    File and prompt content loss fixed in rewind and permission approvals

    Fixed an issue where `/rewind` could restore an older file version after two queued prompts or consecutive skill runs. Text and images added while approving a permission prompt are no longer lost if the command or tool later fails.

    Usage details

    Who can use it: Claude Code’s `/rewind` and permission prompt flows.

    How to use it: Update to 2.1.330; use `/rewind` after changing a file, or add text/images to a permission prompt and check what happens after a failed command.

    Tip: Save a Git diff before rewinding critical files.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APISelf-hosted runner

    Two Prometheus metrics added for runner job order and first session

    Two metrics were added to the self-hosted runner `/metrics` endpoint to measure the time from job order to registration acceptance and first session delivery.

    Usage details

    Who can use it: Self-hosted runners launched by an orchestrator.

    How to use it: Collect the `claude_code_self_hosted_runner_order_to_registered_seconds` and `claude_code_self_hosted_runner_order_to_first_session_seconds` metrics from the `/metrics` output.

    Tip: Track the two metrics separately to distinguish queue/startup latency from session delivery latency.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • SettingClaude Code

    Unnecessary file diffs are skipped by default in self-hosted Bash output

    Self-hosted runner Bash commands no longer produce file diffs when there are no changes, so diffs may be absent from transcripts. If needed, managed settings can force every diff to be included.

    Usage details

    Who can use it: Self-hosted runner sessions.

    How to use it: If you need a diff for every command, set `bashEditDiffEnabled: true` in managed settings.

    Tip: Keep the default to reduce transcript size; explicitly enable the setting in audit workflows.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APISelf-hosted runner

    Choose a startup directory within the repository for self-hosted runners

    The runner can now start in a specific directory within the repository. Settings and hooks continue to load as if no startup directory had been specified.

    Usage details

    Who can use it: Organizations using self-hosted runners.

    How to use it: Pass `--start-dir` when starting the runner, or set the `SELF_HOSTED_RUNNER_START_DIR` environment variable to a directory within the repository.

    Tip: In monorepos, narrow the working directory, but test that shared settings and hook files in parent directories are still picked up.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AvailabilitySelf-hosted runner

    Self-hosted runner reports restart and queue errors more clearly

    Fixed issues where the runner ignored a session after a failed report, a slow spawn hook locked up free slots, a full disk accepted a session, and an incorrect provider/credential error was shown. The status of pushed work and missing repositories in resumed sessions is also explained more clearly.

    Usage details

    Who can use it: Organizations using self-hosted runners.

    How to use it: Upgrade the runner to 2.1.330; check logs for the new explanations in full-disk, proxy 403, and repository credential scenarios.

    Tip: Track runner metrics alongside startup logs to distinguish infrastructure issues from repository issues.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AgentClaude Code

    The skill model setting no longer changes subflows in the main session

    Skills with a `model` setting and no `context: fork` now leave subagents, forks, compaction, and the `/recap` flow on the session’s model.

    Usage details

    Who can use it: Claude Code skills with a model setting.

    How to use it: Keep the `model` field in the skill definition; if a subflow needs to run on a separate model, explicitly design for `context: fork` behavior.

    Tip: When updating skills, test the main model, compaction model, and subagent model separately.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APISelf-hosted runner

    spawn-runner hooks can now run concurrently with limits

    `spawn-runner` hooks can run without waiting for previous hooks to finish; the default maximum is four per orchestrator. The limit can be lowered to one for incompatible hooks.

    Usage details

    Who can use it: Self-hosted runner orchestrators.

    How to use it: If needed, start the orchestrator with `--hook-concurrency 1`; with the default value of four, verify that hooks can safely run concurrently.

    Tip: If hooks use shared files or ports, run them one at a time first, then increase concurrency.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Teleport no longer runs branch-switching commands itself in some setups

    In some setups, `claude --teleport` prints the Git commands required to switch branches instead of running them automatically. When `-p` is used, it returns exit code 1 until the branch is switched.

    Usage details

    Who can use it: Setups that use Teleport, especially automated or headless workflows.

    How to use it: Run the Git commands shown in the Teleport output, switch to the correct branch, and then restart the session; check the exit code in `-p` workflows.

    Tip: Make branch switching an explicit step in CI scripts and evaluate Claude Code’s exit code separately.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Desktop

    Third-party Desktop sessions recover better after large PDFs

    In some third-party provider sessions outside Scheduled tasks, a PDF that exceeded the context limit after a long conversation could permanently end the turn with Prompt is too long. This has been fixed. An order dependency when choosing between user and project agent files with the same name has also been resolved.

    Usage details

    Who can use it: Claude Desktop sessions using third-party providers and projects with agent files that have the same name.

    How to use it: Update Desktop to 2.1.330; add a large PDF after a long conversation and test that the session continues.

    Tip: Providing large PDFs at the start of a long context or in sections reduces the risk of hitting provider limits.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code for VS Code

    VS Code prompt, diff, question, and keyboard workflows are more reliable

    Fixed issues where VS Code’s attached terminal opened and took focus with every prompt, a line break accidentally ran a command while copying code, an old answer was submitted in question cards, and Accept/Reject buttons appeared on the wrong diff.

    Usage details

    Who can use it: Claude Code for VS Code.

    How to use it: Update the extension to 2.1.330; retest question cards, proposed diffs, code copying, and attached terminal workflows.

    Tip: Before production use, also test question and diff workflows with team members who use keybindings and IMEs.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • SettingClaude Code for VS Code

    Escape, pinned prompt, and context indicator settings added to VS Code

    The `claudeCode.escapeStopsClaude`, `claudeCode.pinPrompt`, and `claudeCode.contextUsageThreshold` settings were added. They control Escape’s stop behavior, whether the prompt stays pinned, and the context indicator’s threshold.

    Usage details

    Who can use it: Claude Code for VS Code.

    How to use it: Search for these three `claudeCode.*` settings in VS Code Settings; turn off `claudeCode.escapeStopsClaude` if you don’t want Escape to stop Claude.

    Tip: In long conversations, set `contextUsageThreshold` low to get an early warning.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • FeatureClaude Code for VS Code

    Added /fork and /workflows commands in VS Code

    The VS Code integration now includes `/fork` to branch a conversation after the latest response and `/workflows` to open running workflows in the agent map.

    Usage details

    Who can use it: The Claude Code VS Code extension.

    How to use it: Type `/fork` or `/workflows` in the Claude panel in VS Code; for a fork, run the command after the conversation’s latest response.

    Tip: Use `/fork` to isolate experimental changes and keep the main conversation’s context clean.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    WebFetch long waits and empty results fixed

    Issues that caused WebFetch to lock up sessions for minutes or return empty results on some pages have been fixed. Page summaries now track the `/effort` level more accurately, even if the gateway rejects the small fast model.

    Usage details

    Who can use it: Sessions using WebFetch in Claude Code.

    How to use it: Update Claude Code to 2.1.330 and retry the same URLs with WebFetch; if the issue persists, check the `/effort` setting.

    Tip: On very large pages, specify the target section explicitly to reduce unnecessary waiting.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • AvailabilitySelf-hosted runner

    If a checkout hook returns 108, the runner stops without trying another runner

    When a checkout hook exits with code 108, the self-hosted runner immediately stops session startup and does not switch to another runner.

    Usage details

    Who can use it: Self-hosted runner checkout hooks.

    How to use it: Document that exit code 108 means an intentional stop in your checkout hooks, and update session startup monitoring rules accordingly.

    Tip: Don’t use 108 for transient errors; define a separate exit path for cases where a retry is expected.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Config search, local statistics, and Ultrareview feedback improved

    `/config` search prioritizes setting-name matches; `/resume` no longer lists sessions that were only opened and exited. `/stats` now shows the local time zone, while `/ultrareview` more accurately shows incorrect repository status and the Team/Enterprise spending share.

    Usage details

    Who can use it: Claude Code terminal interface; Team and Enterprise plans for Ultrareview.

    How to use it: Search for a setting name in `/config`, or run `/resume`, `/stats`, or `/ultrareview`; check the new descriptions and local dates.

    Tip: Don’t start Ultrareview before checking the remaining monthly limit for cost approval.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    design-sync now builds its reference Storybook only from the project installation

    The fallback build for the `/design-sync` reference Storybook now uses only the Storybook installed in the project; it no longer looks for an external installation.

    Usage details

    Who can use it: Projects using Claude Design and `/design-sync`.

    How to use it: Add the Storybook dependency to the project and verify the project installation before running `/design-sync`.

    Tip: Pin the Storybook version used in CI with the lockfile.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationClaude Code

    A clear warning added for marketplaces not in the session repository

    Cloud and self-hosted runner sessions now explain why they couldn’t be retrieved when the plugin marketplace repository isn’t among the session repositories.

    Usage details

    Who can use it: Plugin marketplace use in Cloud and self-hosted runner sessions.

    How to use it: When the warning appears, add the marketplace repository to the list of repositories accessible to the session, or move the marketplace to a suitable repository.

    Tip: For automated installations, verify marketplace repository access before starting the session.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • Plugin / IntegrationSelf-hosted runner

    Plugin name and next steps shown when a plugin repository is rejected

    When Anthropic-managed git rejects a plugin’s own repository, the self-hosted runner now names the plugin and explains what to do.

    Usage details

    Who can use it: Self-hosted runner sessions using Anthropic-managed git.

    How to use it: Follow the plugin name and recommended repository handling path in the warning; if necessary, exclude the plugin from the session’s repository scope.

    Tip: If the plugin repository and working repository are the same, plan the installation through a separate repository.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • APIClaude Code

    Prompt hook model field accepts haiku and sonnet aliases

    The `model` field in prompt hook configuration now accepts model aliases such as `haiku` and `sonnet`.

    Usage details

    Who can use it: Claude Code configurations using prompt hooks.

    How to use it: Enter the supported `haiku` or `sonnet` alias in the `model` field of the prompt hook definition, then test the hook.

    Tip: In production automation, using a fixed model ID instead of an alias can reduce the impact of model changes.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code

    Unrecognized Sonnet and Haiku names use a shorter system prompt

    Unrecognized model names containing `sonnet` or `haiku` now use a shorter system prompt, like other unrecognized names.

    Usage details

    Who can use it: Claude Code sessions using unrecognized model names through a custom gateway or provider.

    How to use it: Set your custom model name and start the session; use `CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0` if you need the full system prompt for all unrecognized names.

    Tip: Standardize model naming in the gateway to reduce surprises in prompt size and model behavior.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları
  • CapabilityClaude Code for VS Code

    Prompt and session group actions expanded in the VS Code sidebar

    The Open in Side Bar command now accepts a `prompt` argument for keybindings and extensions; Move up and Move down actions were added to the session group menu.

    Usage details

    Who can use it: The Claude Code for VS Code extension.

    How to use it: Pass the `prompt` argument to the Open in Side Bar command in a keybinding definition; select Move up or Move down from the session group menu.

    Tip: Reduce repetitive context entry by opening frequently used prompts with a keybinding.

    ✓ Verified by official sourceSource: Claude Code 2.1.330 sürüm notları

OpenAI · ChatGPT

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

New in this edition · 6

  • FeatureChatGPT Space

    ChatGPT Space begins rolling out globally to Plus users

    ChatGPT Space is now available to Plus users; users can collaborate with ChatGPT and teammates on the same page, presentation, and site. Pages and slides can update automatically with information from connected apps and tools.

    Usage details

    Who can use it: Global rollout has begun for Plus users aged 18 and older. Plus, Pro, Business, and Enterprise accounts with access to Space can open shared pages or presentations; Free users cannot. Available on web, desktop, iOS, and Android; there are no public links.

    How to use it: Open Space in the ChatGPT sidebar, where it replaces Library. Choose a template or create a page from an existing conversation; add your files and use Share to grant viewing or editing permission.

    Tip: For pages that update automatically from connected apps, also review sharing permissions and any sensitive information copied to the page.

  • SettingChatGPT Enterprise, Edu, Healthcare ve Business

    Per-person SSO exemptions introduced for external workspace members

    Workspace owners can grant per-person exemptions from mandatory SSO to external members whose email domain has not been verified by the workspace. New external members follow the workspace’s SSO policy by default; exemptions can be granted to preserve existing members’ access.

    Usage details

    Who can use it: Applies to ChatGPT Enterprise, Edu, Healthcare, and Business workspaces. Affects only the SSO requirement for the relevant workspace; owners can change the setting, while owners and admins can review it.

    How to use it: In Admin Console, select the relevant ChatGPT workspace and open the member from the Members section. Under Workspace login, select Allow login without SSO or Follow workspace SSO policy for Single sign-on (SSO), click Save changes, and verify the setting by refreshing.

    Tip: During security reviews, check external members individually and use Follow workspace SSO policy by default; removing and re-inviting a member deletes the existing exemption and does not restore it.

  • ModelOpenAI API Live

    GPT-Live-1.1 launches with multilingual audio improvements and WebRTC standby

    The OpenAI API now offers the gpt-live-1.1 model, with more accurate transcription, more consistent accents, and improved audio pronunciation in multilingual conversations. GPT-Live-1 and GPT-Live-1.1 WebRTC sessions support standby mode, which pauses inference during inactivity and resumes it when speech begins.

    Usage details

    Who can use it: Available on v1/live/sessions; the model accepts audio and text and generates audio and text. GPT-Live-1.1 is not supported in Chat Completions or Responses. Active Live time costs 0,05 US dollars per minute, and standby time costs 0,02 US dollars; the two models have separate concurrent session limits.

    How to use it: Use gpt-live-1.1 as the model in a Live session. In your WebRTC configuration, set session.wake_policy.type to speech and session.idle_timeout_s to a value between 5–60 seconds; keep the microphone track open during standby.

    Tip: Use standby for long periods of silence; to reduce false wake-ups, raise session.wake_policy.vad_threshold above the default of 0,5 if needed.

  • FeatureChatGPT mobil uygulaması

    New ChatGPT mobile app experience launches with Home and Explore tabs

    The redesigned ChatGPT mobile app provides faster access to favorite tools through Home and Explore tabs. Dot moves to its own tab; users can pin tools such as Codex, Projects, and Images to the tab bar.

    Usage details

    Who can use it: Rolling out first to Pro users aged 18 and older in regions where Dot is available, as well as a small eligible group of Plus users. The European Economic Area, the United Kingdom, Switzerland, and Korea are excluded; the rollout is for the latest iPhone and Android apps. iPhone SE, iPhone Duo, and iPad are not yet supported.

    How to use it: Update the app from the App Store or Google Play. From Home, open Settings using the profile icon and enable New ChatGPT design if it appears; in Explore, press and hold a tool and select Pin to tab bar.

    Tip: Pin a few of the tools you use most. If the new design does not appear, your account or device may not yet be included in the rollout, and you can continue using the classic experience.

  • APIOpenAI Node SDK

    gpt-live-1.1 added to OpenAI Node SDK Live model options

    OpenAI Node SDK v7.61.0 adds support for gpt-live-1.1 to the Live model options. Node.js applications can now select the new multilingual audio model in Live sessions through the SDK types.

    Usage details

    Who can use it: Node.js applications using openai-node v7.61.0 and organizations with Live API access.

    How to use it: Update the SDK with the npm install openai@7.61.0 command and set gpt-live-1.1 as the model in your Live session configuration.

    Tip: Validate the model ID against the Live model types in the SDK version you use rather than spreading it as a hard-coded string; also check your API organization limits.

    ✓ Verified by official sourceModels: gpt-live-1.1Source: OpenAI Node SDK sürümleri
  • APIOpenAI Node SDK

    OpenAI Node SDK adds Live WebRTC standby support

    OpenAI Node SDK v7.62.0 adds support for keeping a Live WebRTC connection open while pausing inference during inactivity and resuming it when speech begins. This provides lower standby costs and preserves the same session context for GPT-Live-1 and GPT-Live-1.1 in Node.js applications.

    Usage details

    Who can use it: openai-node v7.62.0 and Live WebRTC sessions; applies to GPT-Live-1 and GPT-Live-1.1.

    How to use it: Update the SDK with the npm install openai@7.62.0 command. In a WebRTC Live session, set session.wake_policy.type to speech and session.idle_timeout_s to 5–60 seconds; listen for the session.paused and session.resumed events.

    Tip: Choose idle_timeout_s based on actual gaps in speech; values that are too short can cause unnecessary wake-ups, while values that are too long can lead to active billing.

    ✓ Verified by official sourceModels: GPT-Live-1, GPT-Live-1.1Source: OpenAI Node SDK sürümleriSource: Live WebRTC standby yapılandırma kılavuzu

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 · 10

  • CapabilityGemini Enterprise Agent Platform

    Supervised fine-tuning and checkpoints for Gemini 3.8 Flash reach GA

    Supervised fine-tuning, continuous tuning, and tuning checkpoints for Gemini 3.8 Flash are now generally available. Tuning jobs can run on us and eu multi-region endpoints; tuned models can be served on us, eu, or global endpoints.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform; us/eu for tuning and us/eu/global multi-region endpoints for serving tuned models.

    How to use it: In the Google Cloud console, go to Gemini Enterprise Agent Platform > Models/Tuning > Supervised fine-tuning and select Gemini 3.8 Flash. Specify the training and serving endpoints, then start the tuning job.

    Tip: Create a fixed evaluation set before tuning; measure the tuned model against the base Gemini 3.8 Flash for quality, latency, and cost.

    ✓ Verified by official sourceModels: Gemini 3.8 FlashSource: Google Cloud release notes — Gemini 3.8 Flash tuning GA
  • Plugin / IntegrationGemini Enterprise

    Third-party data stores reach GA in Gemini Enterprise

    Gemini Enterprise has made many third-party data stores on its official list generally available, including AirOps, Airtable, Asana, GitLab, Linear, Stripe, Zapier, and Zoho Projects. These connectors can now be used in enterprise search and action workflows with GA status instead of preview.

    Usage details

    Who can use it: All third-party data stores declared GA on Gemini Enterprise’s official November 20 list, including AirOps, Airtable, Asana, GitLab, Linear, Stripe, Zapier, and Zoho Projects.

    How to use it: In the Gemini Enterprise admin console, go to Data stores/Connectors > Add data store, select the GA connector you need, and complete the setup and authorization steps.

    Tip: Enable only the data stores you actually need, and configure each connector’s access scope separately with the least privilege.

  • AvailabilityWorkspace Studio

    Data Regions support reaches general availability in Workspace Studio

    Workspace Studio now supports policies that require workflow data to remain in the US or Europe data region, both at rest and in use. A dedicated control in the Admin console lets admins restrict out-of-region steps and generative AI capabilities.

    Usage details

    Who can use it: Google Workspace Business Standard/Plus, Enterprise Standard/Plus, Education Standard/Plus, Frontline Plus, and the relevant Assured Controls/Enterprise Data Regions add-ons; processing region coverage varies by edition.

    How to use it: Configure the Workspace Studio control at Admin console > Account > Data Regions > Advanced Settings or Data > Compliance > Data Regions > Advanced Settings.

    Tip: Audit workflows before disabling out-of-region features; doing so may disable Gemini AI steps and some Tasks, Forms, Notebook, webhook, and Drive triggers.

  • FeatureGoogle Vids

    Reusable media assets for Gemini Omni arrive in Google Vids

    Google Vids now stores and reuses the images, videos, and audio used to generate videos with Gemini Omni in a new Assets section within the same project. The feature is on by default, and users can quickly select new video inputs later within the same project.

    Usage details

    Who can use it: Google Vids; Business, Enterprise, Education, Frontline, Nonprofits, Individual, Workspace Individual, and personal Google account users, as well as Google AI Pro/Ultra users; Scheduled Release rollout begins December 2.

    How to use it: In a Google Vids project, start Gemini Omni video generation; open the Ingredients/Assets section and select recently used assets, or add media from Drive, Upload, or Photos.

    Tip: Keep the same brand images and audio in the Assets section to maintain a consistent visual identity and speed up production across consecutive videos.

    ✓ Verified by official sourceModels: Gemini OmniSource: Workspace Updates — Google Vids’te medya yeniden kullanımı
  • OtherGemini CLI

    Gemini CLI nightly fixes `$` in templates and stream stability issues

    The November 20 nightly build preserves literal `$` characters in template placeholders, reduces Windows PowerShell substitution errors, and fixes several hangs in network, pipe, and retry flows. It also includes fixes to prevent PDF and telemetry data from retaining unnecessary memory.

    Usage details

    Who can use it: Gemini CLI nightly channel; build v0.71.0-nightly.20261120.gd4cb1713f.

    How to use it: Nightly users should upgrade Gemini CLI to the v0.71.0-nightly.20261120.gd4cb1713f GitHub tag and retest template, PowerShell, and headless flows.

    Tip: Keep a rollback tag before pinning the nightly build in CI; this channel delivers new fixes early but carries a risk of regressions.

  • Plugin / IntegrationGemini Enterprise

    Microsoft 365 data store reaches GA in Gemini Enterprise

    Gemini Enterprise has made generally available access to SharePoint, OneDrive, Excel, Outlook, calendars, contacts, tasks, Teams, and Microsoft Entra data through a single Microsoft 365 data store. Users can search, read, create, and update content using actions enabled by an admin.

    Usage details

    Who can use it: Gemini Enterprise; GCC tenants are supported with Google-managed OAuth or their own Microsoft Entra app, while GCC High and DoD tenants are supported with their own Entra app.

    How to use it: In the Gemini Enterprise admin console, go to Data stores > Add data store > Microsoft 365; choose Google-managed OAuth or your own Microsoft Entra app and grant the required Microsoft Graph permissions.

    Tip: Enable only the actions you need to keep Microsoft Graph permissions and the operations Gemini can perform within the narrowest scope.

  • AvailabilityGemini CLI

    Gemini CLI 0.70.0 moves from preview to stable

    Gemini CLI v0.70.0 was released on the stable channel after the previously published v0.70.0-preview.0. This is the stable distribution of the preview release’s features and fixes for production use.

    Usage details

    Who can use it: Gemini CLI stable channel.

    How to use it: Install the v0.70.0 tag from the official GitHub Releases page and update your existing Gemini CLI installation to this version.

    Tip: Pin the stable release for production and CI work; use preview and nightly releases in separate test environments.

    ✓ Verified by official sourceSource: Gemini CLI v0.70.0 GitHub sürümü
  • FeatureGoogle Search

    Collapsible sidebar for Gemini chats launches on the Google Search homepage

    Google Search VP Robby Stein announced a globally rolling-out collapsible sidebar on the desktop homepage; the panel makes it easier to return to recent AI searches. 9to5Google and Android Authority reported that the panel adds Gemini chat/history, tasks and new chat shortcuts, along with buttons for image generation, file analysis and brainstorming; these details are not included in a separate Google release note.

    Usage details

    Who can use it: Global phased rollout on the desktop Google Search homepage, especially in AI Mode; it may not appear for everyone at the same time.

    How to use it: Open AI Mode in desktop Google Search; click the AI/Gemini panel on the left edge of the interface to open the collapsible sidebar.

    Tip: Return to ongoing research from the sidebar; use the central search box for a one-off regular web search.

  • SettingWorkspace Studio

    A Collaboration steps management setting is now available for Workspace Studio flows

    The new Collaboration steps setting lets administrators control whether Workspace Studio flows can share information with others or modify shared resources. The setting is on by default; administrators can limit collaborative actions at the domain, organizational unit or group level.

    Usage details

    Who can use it: Google Workspace Business Starter/Standard/Plus, Enterprise Standard/Plus, Education Fundamentals/Standard/Plus, Google AI Pro for Education, Teaching and Learning and AI Expanded Access.

    How to use it: Go to Google Admin console > Workspace Studio > Collaboration & approvals and change the Collaboration steps setting for a domain, organizational unit or group.

    Tip: Test it with a small group first; enable flows’ permissions to send email, post in Chat and edit shared files only as broadly as needed.

  • OtherGemini CLI

    Gemini CLI 0.71.0-preview.0 fixes a hang when handling PowerShell

    Gemini CLI v0.71.0-preview.0 parses PowerShell commands asynchronously, preventing the spinning loading indicator in the user interface from freezing. This is the current test release on the preview channel.

    Usage details

    Who can use it: Gemini CLI preview channel; especially Windows PowerShell command flows.

    How to use it: Update to the v0.71.0-preview.0 tag from the preview channel and test PowerShell commands with this version.

    Tip: Before adding the preview release to CI, verify PowerShell-based tasks and cancellation/retry behavior in a separate test repository.

Meta · Meta AI

Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 3

  • ModelUMA Nitro

    Meta releases faster UMA Nitro model for atomistic inference

    Meta has released UMA Nitro, a faster version of Universal Model for Atoms. Nitro uses fused Triton and CuTe kernels for supported single-system energy, force, and stress inference.

    Usage details

    Who can use it: Researchers and developers with UMA access through FAIR Chemistry and Hugging Face; Nitro’s default fast path runs on CUDA compute capability 8.x–12.x, meaning Ampere, Hopper, and Blackwell GPUs. Unsupported hardware or workloads fall back to other inference settings.

    How to use it: Install the fairchem-core package, request access to facebook/UMA on Hugging Face, and log in with huggingface-cli login. In Python, use the call pretrained_mlip.get_predict_unit("uma-s-1p2p1", device="cuda"); the default inference setting is nitro.

    Tip: If precision is important for higher-order derivatives such as Hessians or phonons, use inference_settings="high-precision"; the documentation states that Nitro uses TF32 on Hopper and Blackwell.

  • AvailabilityUMA Playground

    UMA Playground opens to high school chemistry teachers

    According to Meta’s November 20 announcement, UMA Playground offers an experience for building molecules, watching bonds form and break, and exploring chemical reactions. Discovery Education has made the tool available to high school chemistry teachers alongside classroom guides.

    Usage details

    Who can use it: High school chemistry teachers through Discovery Education; Meta’s UMA Playground demo is also available via a public link.

    How to use it: Open UMA Playground at https://aidemos.atmeta.com/uma?view=playground; for classroom use, follow the relevant classroom guides from Discovery Education.

    Tip: Prepare examples of bonds forming and breaking before class, then have students modify the same molecules to visually compare reactions.

  • APIWhatsApp Marketing Messages API

    New error code announced for undeliverable messages in WhatsApp Marketing Messages API

    Meta announced that starting in February 2027, it will return error code 131066 instead of 131049 when marketing messages are not delivered because of template quality or max price. 131049 will continue to indicate other undeliverable message statuses to support a healthy messaging experience.

    Usage details

    Who can use it: Developers using WhatsApp Marketing Messages API; the behavior change will begin in February 2027.

    How to use it: Update your message-sending error-handling code to treat 131066 as non-delivery due to template quality or price; continue supporting 131049 during the transition.

    Tip: Do not combine 131066 and 131049 into a single failure category; use separate remediation and monitoring flows for quality issues and max price issues.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 10

  • CapabilityKimi Code

    Draft effort inherited from the global thinking setting is preserved

    A separate fix was made for the draft `thinking effort` value inherited from the global thinking setting reverting to the model default after a message is sent. This was explicitly released in the 1.4.2 and CLI 2.7.2 patch.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.2 and Kimi Code CLI 2.7.2; Web users who use the global thinking setting.

    How to use it: Set your global thinking setting, write and send a message in a new draft; the effort value should not reset to the model default after sending.

    Tip: When switching between models, keep the global setting fixed and open a new draft; this makes it easier to track the difference between model-specific defaults and your global preference.

  • AvailabilityKimi K3

    Kimi K3 model card points to Kimi K3.1 Pro

    `new_version: moonshotai/Kimi-K3.1-Pro` was added to the metadata of Moonshot AI’s `Kimi K3` model card on Hugging Face. Hugging Face now shows visitors to the old card a link to switch to the newer Kimi K3.1 Pro model; there is no claim that the weights have changed.

    Usage details

    Who can use it: All users who visit the `moonshotai/Kimi-K3` model card on Hugging Face.

    How to use it: Open the `moonshotai/Kimi-K3` model card and switch to the `moonshotai/Kimi-K3.1-Pro` card using the newer-version notice shown there.

    Tip: When starting a new deployment or evaluation, follow the redirect and use the latest usage instructions on the Kimi K3.1 Pro card.

  • CapabilityKimi Code CLI

    Queued prompts wait until the background notification turn finishes

    Prompts queued while background notifications are producing a response turn are no longer processed prematurely; they remain pending until the relevant notification turn is complete. This prevents queued messages from running out of order.

    Usage details

    Who can use it: Kimi Code CLI 2.7.1 and users on the same Web flow.

    How to use it: Send a new message while a background task’s notification turn is in progress; the message waits in the queue and is processed when the notification turn finishes.

    Tip: Queue follow-up instructions that depend on the background notification’s result so they’re less likely to run with outdated context.

  • CapabilityKimi Code

    Turn-specific forking is fixed in restored sessions after steer

    A bug that caused forking from a specific turn to fail after restoring a session that included a steer has been fixed. This makes it more reliable to start independent work from a past turn.

    Usage details

    Who can use it: Users restoring sessions with Kimi Code Desktop 1.4.1 and CLI 2.7.1.

    How to use it: Restore a session containing a steer, open the relevant past turn, and select the `Fork` or `Fork from here` action.

    Tip: Before an important branch, reload the session and verify the fork; this reveals early whether the steer history has been preserved.

  • CapabilityKimi Code

    Draft thinking effort selection no longer resets to the default after sending

    A bug that caused the `thinking effort` selected for a draft in Kimi Code Web and Desktop to revert to the model default when the message was sent has been fixed. The same fix was also applied to the Web flow in CLI 2.7.1.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.1, CLI 2.7.1, and the Web interface provided with them.

    How to use it: Open a new draft, select a `thinking effort` for the model, and send the message; your selection will be preserved for the next request.

    Tip: For long tasks, check the draft effort value before sending the message; your selection should no longer silently revert to the model default.

  • CapabilityKimi Code CLI

    The Web input composer no longer disappears when accessed over insecure HTTP

    A bug that caused the input composer to become invisible when accessing the Kimi Code Web interface over insecure HTTP has been fixed. This keeps message entry available in local or proxy-backed HTTP setups.

    Usage details

    Who can use it: The Web interface provided with Kimi Code CLI 2.7.1; users accessing it over HTTP.

    How to use it: Open the Kimi Code Web interface using an HTTP address and use the input field; the composer will remain visible automatically.

    Tip: Continue to use HTTPS in production; this fix does not eliminate the risks of an insecure connection.

  • AvailabilityKimi K2.5

    The Kimi K2.5 model card points to Kimi K2.6

    `new_version: moonshotai/Kimi-K2.6` was added to the metadata of Moonshot AI’s `Kimi K2.5` model card on Hugging Face. Hugging Face now offers visitors to the Kimi K2.5 card a link to switch to the newer Kimi K2.6 model; this is a change to the card metadata and was not announced as a weights update.

    Usage details

    Who can use it: All users visiting the `moonshotai/Kimi-K2.5` model card on Hugging Face.

    How to use it: Open the `moonshotai/Kimi-K2.5` model card and follow the `moonshotai/Kimi-K2.6` link in the newer-version notice.

    Tip: For new deployments, follow the deployment and API instructions on the Kimi K2.6 card; use the older K2.5 card only for backward compatibility.

  • FeatureKimi Code Desktop

    The current folder can be added using the primary button in the folder picker

    If no subfolder is selected in the folder picker, the current folder can now be added directly to the project using the primary button. There’s no longer a need to use a separate `Select this folder` button.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.1.

    How to use it: In the new-project flow, open the folder picker, stay in the current folder without selecting a subfolder, and click the primary selection button.

    Tip: If you want to use the existing project folder, use the primary button without opening a subfolder; this reduces the chance of selecting the wrong root directory.

  • CapabilityKimi Code Desktop

    The first prompt is preserved during project creation and sign-in prompts

    The first prompt a user writes before proceeding to a project creation or sign-in prompt is no longer lost during these flows. Once the process is complete, they can continue without rewriting the prompt.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.1.

    How to use it: Start a new project, write the first prompt, and complete the project creation or sign-in prompt; check that the prompt remains in the draft.

    Tip: Don’t send long initial prompts before signing in; complete the flow first, then review the preserved draft one last time.

  • FeatureKimi Code CLI

    Kimi Code CLI can start the local server without opening a browser

    With Kimi Code CLI 2.7.0, the local server can be started without opening a browser using the new `kimi server` command. The `--open` option remains available for users who want to open the web interface automatically.

    Usage details

    Who can use it: Kimi Code CLI 2.7.0 and later; all CLI users running the local server.

    How to use it: Run `kimi server` in the terminal; use `kimi server --open` to also open the local web interface.

    Tip: Use the command without `--open` in CI, SSH, or remote-access scenarios to avoid opening an unnecessary browser.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.7.0 GitHub sürümü

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 64

  • FeatureQwen Code

    Qwen Code runs background workflows in headless mode

    The CLI supports running background workflows in headless mode. This enables automated task execution without an interactive terminal open.

    Usage details

    Who can use it: Qwen Code v0.28.3; CLI headless environments.

    How to use it: Start the workflow with the headless execution option and monitor the output in the workflow or session logs.

    Tip: For headless jobs, redirect stdout and error output to a file and monitor the exit code.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • FeatureQwen Code

    Qwen Code CLI adds native screen-reader presentation

    An opt-in native screen-reader view has been added for the CLI. It provides keyboard navigation, announcements of completed responses, and semantic history; Ink remains the default interface.

    Usage details

    Who can use it: Qwen Code v0.28.3; supported terminal screen readers.

    How to use it: Enable native screen-reader presentation in the CLI accessibility settings and start the session with your terminal screen reader.

    Tip: If you don’t use a screen reader, keep the default Ink view and enable this mode only when needed.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • APIQwen Code

    Qwen Code expanded session access for create_sub_session

    create_sub_session can now access any session; in YOLO mode, it can skip the workspace prompt and remove the thread icon.

    Usage details

    Who can use it: Qwen Code v0.28.3; serve API and host integrations in YOLO mode.

    How to use it: Specify the target session when calling create_sub_session through the serve API; if you use YOLO, there is no need to wait separately for workspace approval.

    Tip: Grant access to any session only to trusted hosts, and verify session IDs.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • AgentQwen Code

    Qwen Code sends watched PR updates privately to the linked session

    When CI fails, a conflict occurs, or the head changes on a watched pull request, a private prompt can be sent to the linked Managed Session. The message wakes the session without appearing as a public PR comment.

    Usage details

    Who can use it: Qwen Code v0.28.3; linked Managed Sessions with PR watch.

    How to use it: Link a Managed Session to a pull request and enable PR watch; receive CI or head changes as private prompts within the session.

    Tip: Use this flow only for events that require human intervention to reduce public comment noise.

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

    Qwen Code can skip models that run out of capacity in the same turn

    An optional setting skips models that return capacity errors for the rest of the turn and retries them on the next turn. This makes the model fallback chain smoother during temporary capacity issues.

    Usage details

    Who can use it: Qwen Code v0.28.3; sessions that use a model fallback chain.

    How to use it: In model availability or fallback settings, enable the opt-in option that skips a model until the end of the turn after a capacity error.

    Tip: Enable this option in environments where you prefer uninterrupted task flow over insisting on a single model.

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

    Qwen Code added a Microsoft Teams channel extension

    The Microsoft Teams channel extension supports personal chats, group mentions, and thread isolation. The createTeamsIngress host webhook setup was also added.

    Usage details

    Who can use it: Qwen Code v0.28.3; installations using the Microsoft Teams extension and host webhook.

    How to use it: Install the Teams extension, configure the createTeamsIngress host webhook, and connect the bot to a personal or group chat.

    Tip: Keep thread isolation enabled to prevent different Teams conversations from mixing into the same session context.

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

    Qwen Code can send local media through QQ bots

    An optional QQ bot setting provides tools for the Agent to send local files, images, videos, and voice messages.

    Usage details

    Who can use it: Qwen Code v0.28.3; QQ bot; feature disabled by default.

    How to use it: Enable the relevant off-by-default option in the QQ bot configuration; allow the Agent to use file or media-sending tools.

    Tip: Enable this capability only in trusted QQ chats and limit the local folders from which files can be sent.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • FeatureQwen Code

    Qwen Code adds a workspace artifact library to the Web UI

    The Managed Agent Web UI now displays workspace artifacts from across sessions in a shared library for browsing, previewing, and downloading.

    Usage details

    Who can use it: Qwen Code v0.28.3; Managed Agent Web UI.

    How to use it: Open the workspace artifact library section in the Managed Agent Web UI; select an artifact to preview or download it.

    Tip: When working across many sessions, check the file name and session source together to avoid downloading the wrong artifact.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • SettingQwen Code

    Qwen Code adds a --timeout option for /model --vision

    The Vision bridge timeout can be set for the /model --vision command using the --timeout option. The value can be specified in seconds or ms and takes effect immediately.

    Usage details

    Who can use it: Qwen Code v0.28.3; sessions using the CLI Vision bridge.

    How to use it: Run `/model --vision --timeout 30s` in the CLI, or use a value in ms format; the new timeout takes effect immediately.

    Tip: Increase the timeout for remote or intensive image processing, and keep it short in workflows that need quick feedback.

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

    Qwen Code ends the tool turn when a queued message is sent in ACP

    Ending the tool-calling turn when queued input is sent in an ACP session is now supported. This prevents new user input from getting stuck at the end of an old tool turn.

    Usage details

    Who can use it: Qwen Code v0.28.3; ACP clients and sessions.

    How to use it: In an ACP session, queue new input while the Agent is calling a tool, then send it; wait for the current tool turn to end.

    Tip: For urgent handoffs, let users know how queued input will end the current operation.

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

    Qwen Code preserves error content reported by tools

    An issue that caused reported tool failure content to be lost has been fixed. The Agent can continue to see the cause of the error and tool output more accurately.

    Usage details

    Who can use it: Qwen Code v0.28.3; tool call failures.

    How to use it: When a tool fails, open the error content in the response; use the same details to fix the task and run it again.

    Tip: Before clearing error content, note the root cause and file path information.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • FeatureQwen Code

    Qwen Code moved the artifact list to the Web Shell right panel

    The Artifacts (N) entry in the header has become a right-panel entry that opens a tab with the session-specific artifact list when clicked.

    Usage details

    Who can use it: Qwen Code v0.28.3; Web Shell sessions and tasks that generate artifacts.

    How to use it: In a Web Shell session, click Artifacts in the right panel to open the tab with the session’s list.

    Tip: Check the artifact count quickly in the header, then open only the list for the relevant session.

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

    Qwen Code added a daemon restart recovery inbox

    A recovery inbox was added for situations that require recovery after a daemon restart. This makes it easier to find pending tasks or states after a restart.

    Usage details

    Who can use it: Qwen Code v0.28.3; qwen serve and Web Shell daemon sessions.

    How to use it: After restarting the daemon, open the recovery inbox and check pending recovery records one by one.

    Tip: Before clearing recovery records, verify that the relevant session has actually reconnected.

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

    Qwen Code turned DingTalk replies into editable cards

    Managed Agent replies in DingTalk are now shown as a single editable native card that updates as chunks arrive; the same card is refreshed instead of showing plain-text chunks.

    Usage details

    Who can use it: Qwen Code v0.28.3; DingTalk Managed Agent channel.

    How to use it: Connect the DingTalk channel to Managed Agent and send a task; open the reply card and follow the updated content.

    Tip: Copy long replies once the card has reached its final state.

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

    Qwen Code can route incoming injected messages to a running turn

    Messages from the inject channel can now be routed to a running turn; the Agent can read them at the boundary of the next tool batch.

    Usage details

    Who can use it: Qwen Code v0.28.3; Managed sessions using the inject channel.

    How to use it: Configure the inject channel and send a message to the running session; expect the Agent to process it at the next tool-batch boundary.

    Tip: Keep urgent instructions short; long messages can unnecessarily expand the running turn’s context.

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

    Qwen Code added support for Gemini proxy streaming and model refresh

    Support was added for streaming through a Gemini proxy and refreshing the model catalog. Retrieving the current model list is now more reliable for proxy-based connections.

    Usage details

    Who can use it: Qwen Code v0.28.3; Gemini proxy connections.

    How to use it: Configure the Gemini proxy connection, refresh the model catalog, and run a streaming request.

    Tip: After refreshing the models, check that the selected model ID matches the proxy’s current name.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • AgentQwen Code

    Qwen Code added support for external Shell workers in Managed Agent

    Managed Agent supports external Shell workers publishing control-plane events. This enables local or separate execution workers to report status to the Managed flow.

    Usage details

    Who can use it: Qwen Code v0.28.3; Managed Agent external Shell worker profile.

    How to use it: Configure the external Shell worker profile and allow the worker to publish control-plane events; monitor Managed session status in the Web UI.

    Tip: Limit worker permissions to only the required workspace and event scope.

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

    Qwen Code displays reply lines rewritten by Mod in Web Shell

    Assistant reply lines rewritten by Mod are now visible in Web Shell. The original text is preserved for history, copying, search, and branching.

    Usage details

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

    How to use it: Enable the Mod rewrite flow; compare the displayed line in Web Shell with the original text preserved in the history.

    Tip: Remember that the original text is preserved when searching or branching.

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

    Qwen Code accepts voice attachments from QQ bots via CDN

    The QQ bot can accept voice attachments from QQ’s voice CDN. This makes flows that receive voice content through QQ more reliable to process.

    Usage details

    Who can use it: Qwen Code v0.28.3; QQ bot integration.

    How to use it: Configure the QQ bot channel and send a message with a voice attachment; the bot receives the attachment as session input.

    Tip: Make sure the QQ CDN link for the voice attachment is accessible to the bot.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • FeatureQwen Code

    Qwen Code shows workflow phases before they start

    The workflow’s meta.phases definitions appear as Not started lanes before it begins. A lane updates when the script reaches the corresponding phase.

    Usage details

    Who can use it: Qwen Code v0.28.3; Web Shell workflow views.

    How to use it: Run a workflow with defined phases and open the workflow view in Web Shell to see the phases before they start.

    Tip: In long workflows, keep phase names short and action-oriented to make status tracking easier.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • SettingQwen Code

    Qwen Code added a setting to preserve MCP connections in suspended sessions

    When the qwen.managed-agent.runtime-broker.mcp-parking.enabled setting is enabled, idle MCP connections are retained during suspension. A reopened session can reuse the same connections.

    Usage details

    Who can use it: Qwen Code v0.28.3; sessions using Managed Agent Runtime Broker; opt-in.

    How to use it: Add `qwen.managed-agent.runtime-broker.mcp-parking.enabled` with a value of true to the configuration, then reopen the Managed session.

    Tip: Use this setting only for MCP services where reopening is costly, to avoid consuming connection resources.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • FeatureQwen Code

    Qwen Code can explain background tasks from composer chips

    A user-facing feature was added to explain background tasks from their composer chips in Web Shell. You can now see what a task is doing from its chip without opening its status.

    Usage details

    Who can use it: Qwen Code v0.28.3; Web Shell composer chips.

    How to use it: Start a background task in Web Shell and open the explain action on its composer chip.

    Tip: When multiple background tasks are running, open each chip separately to track the right task.

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

    Qwen Code reports attachments in unbound Managed sessions

    Support for input attachments in unbound Managed sessions was announced. This lets users find out earlier whether a session accepts attachments.

    Usage details

    Who can use it: Qwen Code v0.28.3; Managed Agent sessions.

    How to use it: Open the Managed session details; check the input attachment capability shown while the session is unbound, then send an attachment.

    Tip: Before sending an attachment, verify the session’s target workspace and permissions.

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

    Qwen Code lets you draft while history is loading

    Fixed an issue that prevented drafting in the composer while session history was loading.

    Usage details

    Who can use it: Qwen Code v0.28.3; Web Shell session history loading screen.

    How to use it: Open the session and type in the composer while history is loading; the draft will be preserved when loading completes.

    Tip: For long histories, use the wait time productively by starting to write your first draft.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • SettingQwen Code

    Qwen Code lets hosts turn off the composer glow

    Web Shell host applications can disable the composer glow visual effect.

    Usage details

    Who can use it: Qwen Code v0.28.3; hosts embedding Web Shell.

    How to use it: Set the composer glow option to false in the host configuration, then reload Web Shell.

    Tip: Turning off this visual effect can simplify the interface on corporate or low-powered devices.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • SettingQwen Code

    Qwen Code exposes the transcript overscan setting to WebShell hosts

    The WebShell transcript component provides an overscan option for setting or disabling the offscreen row buffer in the virtualized message list.

    Usage details

    Who can use it: Qwen Code v0.28.3; host applications embedding WebShell.

    How to use it: Set the transcript component’s `overscan` prop in the WebShell host configuration; set it to 0 to disable the buffer if needed.

    Tip: In large transcripts, a low but nonzero overscan value preserves scrolling smoothness.

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

    Qwen Code identifies which output budget was exceeded when a write is rejected

    When a write operation is rejected because of an output budget, the error now names the budget. This makes it easier to determine which limit needs adjustment.

    Usage details

    Who can use it: Qwen Code v0.28.3; file writes and output limit errors.

    How to use it: If a write is rejected, read the output budget name in the error message and adjust the relevant model or task limit.

    Tip: Before increasing the budget, try reducing unnecessary output and large diffs.

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

    Qwen Code converts local or failed Markdown images into text chips

    Web Shell displays Markdown images with local paths or images that fail to load as text chips instead of leaving them as broken images.

    Usage details

    Who can use it: Qwen Code v0.28.3; Web Shell Markdown responses.

    How to use it: Open a response containing a local or inaccessible image link in Markdown; view the text chip in Web Shell.

    Tip: Check the path in the chip and re-add the image using the correct workspace path.

    ✓ Verified by official sourceSource: Qwen Code Release v0.28.3
  • AvailabilityAlibaba Cloud Model Studio

    mindlab/macaron-v1.1-venti-pro launches in Singapore

    mindlab/macaron-v1.1-venti-pro has been added to the International deployment scope in the Singapore region. The model supports a 1M context window, visual understanding, and Chat, Agent, Coding, and UI4A capabilities; it can be used with OpenAI-compatible and Anthropic-compatible APIs.

    Usage details

    Who can use it: Alibaba Cloud Model Studio Singapore; International deployment scope.

    How to use it: Select the Singapore region in Model Studio and use `mindlab/macaron-v1.1-venti-pro` as the model ID; send a request to an OpenAI- or Anthropic-compatible endpoint.

    Tip: Measure actual token costs on long code or document tasks before filling the 1M context window.

    ✓ Verified by official sourceModels: mindlab/macaron-v1.1-venti-proSource: Model releases and updates
  • AgentQoder CLI

    Qoder CLI now runs Subagent tasks in the background by default

    The default Subagent execution strategy changed from foreground to background. This lets the main session remain interactive while a subtask runs.

    Usage details

    Who can use it: Qoder CLI 1.1.94; TUI sessions using Subagent.

    How to use it: Update Qoder CLI to 1.1.94 and start a Subagent task; it will run in the background by default.

    Tip: If you want to see results immediately for short tasks, explicitly set the Subagent strategy to foreground.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • SettingQoder IDE

    Qoder IDE can restrict enterprise device sign-ins through organization policy

    Enterprise IT administrators can use managed system policies to limit device sign-ins to accounts belonging to specific organizations. This helps prevent unauthorized account use on enterprise devices.

    Usage details

    Who can use it: Qoder IDE 1.41.1; devices subject to enterprise management policies.

    How to use it: Specify the permitted organizations in the managed system policy; when the IDE is reopened, sign-in is limited to accounts from those organizations.

    Tip: Test the allowlist for support accounts and the emergency administrator account before deploying the policy.

    ✓ Verified by official sourceSource: IDE Release Notes
  • SettingQoder CLI

    Qoder CLI sets a 10-minute default timeout for background Bash commands

    The default timeout for Bash commands running in the background is now 10 minutes.

    Usage details

    Who can use it: Qoder CLI 1.1.94; background Bash commands.

    How to use it: Upgrade the CLI to 1.1.94; if needed, set a shorter or longer timeout when running the command or in the settings.

    Tip: For long builds, split the command into steps or set an explicit timeout to avoid exceeding the default.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder IDE

    Qoder IDE speeds up context compaction for some models

    Context compaction takes less time for long chats with some models, reducing wait times during long sessions.

    Usage details

    Who can use it: Qoder IDE 1.41.1; long chats with supported models.

    How to use it: Continue a long chat as usual; compaction is applied automatically as the context limit approaches.

    Tip: For very long sessions, it is still helpful to preserve important decisions in a brief interim summary.

    ✓ Verified by official sourceSource: IDE Release Notes
  • CapabilityQoder IDE

    Qoder IDE reads attachments as needed and evaluates file context

    Instead of reading all attachments at once, the IDE reads them as needed and evaluates whether the open file is relevant to the chat. This is intended to reduce unnecessary context usage.

    Usage details

    Who can use it: Qoder IDE 1.41.1; chats with attachments and open files.

    How to use it: Attach a file to the chat and make a request related to the open file; the IDE adds relevant attachments to the context as needed.

    Tip: Clearly state the file’s role in your request to help the model select the right file.

    ✓ Verified by official sourceSource: IDE Release Notes
  • CapabilityQoder IDE

    Qoder IDE now selects the correct enterprise or official model when they share a name

    Fixed a bug that could select the wrong model when enterprise and official models had the same name. An explicitly selected enterprise model is now preserved when switching models and restoring a chat.

    Usage details

    Who can use it: Qoder IDE 1.41.1; accounts using the Enterprise model catalog.

    How to use it: Explicitly select the enterprise model from the model picker, then check that the selection is preserved after reopening the chat.

    Tip: For models with the same name, verify the provider or Enterprise label before starting a task.

    ✓ Verified by official sourceSource: IDE Release Notes
  • CapabilityQoder IDE

    Qoder IDE stops the Agent from retrying when a file save fails

    Fixed behavior that caused the Agent to continue editing or retrying when a file save failed. The stuck Generating state after file locks and Hook blocks is also recovered.

    Usage details

    Who can use it: Qoder IDE 1.41.1; users editing files with the Agent.

    How to use it: If a file save fails, resolve the error details and restart the request; don’t expect the Agent to continue automatically.

    Tip: For shared files, check that the file isn’t locked by another process before saving.

    ✓ Verified by official sourceSource: IDE Release Notes
  • CapabilityQoder IDE

    Qoder IDE preserves file references in copied messages

    When a sent message is copied from the hover menu and pasted into the input field, its referenced files and attachment labels are preserved. This makes it easier to edit and resend a previous request.

    Usage details

    Who can use it: Qoder IDE 1.41.1; chats with file attachments.

    How to use it: Hover over a sent message, copy it, and paste it into the input field; edit the text and resend it.

    Tip: Avoid unnecessary uploads by checking whether existing references are preserved instead of attaching the files again.

    ✓ Verified by official sourceSource: IDE Release Notes
  • AvailabilityQoder

    Qoder fixed a startup issue after fast Windows updates

    Fixed a bug that prevented Qoder from launching after a fast update in per-user Windows installations.

    Usage details

    Who can use it: Qoder 0.5.17; Windows per-user installations.

    How to use it: Upgrade Qoder to 0.5.17. If the issue persists, close the app and restart it using the regular shortcut.

    Tip: On managed devices, try the repair option after a fast update before manually deleting old installation files.

    ✓ Verified by official sourceSource: Qoder Release Notes
  • CapabilityQoder

    Qoder fixed the app entry point for some marketplace plugins after installation

    Resolved an issue affecting some marketplace plugins that had no app entry point after installation and authorization. These plugins should now be launchable after installation.

    Usage details

    Who can use it: Qoder 0.5.17; desktop users of marketplace plugins.

    How to use it: Install and authorize the plugin from the marketplace. If its entry point doesn’t appear, update or reinstall the plugin.

    Tip: Check the authorization status in the marketplace details before using the plugin.

    ✓ Verified by official sourceSource: Qoder Release Notes
  • CapabilityQoder CLI

    Qoder CLI fixes a crash in the Ctrl+O details view during streaming

    A potential crash that could interrupt the current turn when expanding detailed logs with Ctrl+O during streaming output has been fixed.

    Usage details

    Who can use it: Qoder CLI 1.1.94; streaming TUI output.

    How to use it: Update the CLI to 1.1.94; you can open the details view with Ctrl+O during streaming.

    Tip: During long streams, make sure the cursor is in the active output panel before opening the details view.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder CLI

    Qoder CLI fixes adjacent lines merging after Edit with empty text

    A bug that caused adjacent lines to merge when an Edit operation replaced text with an empty string has been fixed.

    Usage details

    Who can use it: Qoder CLI 1.1.94; sessions using the Edit tool.

    How to use it: Upgrade the CLI to 1.1.94 and use the normal Edit flow to delete empty content.

    Tip: Check the diff after deletion to verify that line breaks are preserved.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder IDE

    Qoder IDE no longer keeps deleted files in the Review list

    Fixed a bug that left files created by the Agent and then deleted in the Review list, locking the review page.

    Usage details

    Who can use it: Qoder IDE 1.41.1; all users of the Review workflow.

    How to use it: Open the Review page. If a deleted file remains, update the IDE and reload the list.

    Tip: Refreshing the working tree before review can help clear out stale diff records.

    ✓ Verified by official sourceSource: IDE Release Notes
  • AvailabilityQoder IDE

    Qoder IDE speeds up authentication after browser sign-in

    The wait time while authentication completes in the IDE after signing in through a browser has been reduced.

    Usage details

    Who can use it: Qoder IDE 1.41.1; accounts using browser-based sign-in.

    How to use it: Sign in to the IDE, complete authorization in the browser, then return to the IDE.

    Tip: If the browser redirect does not complete, wait a few seconds and check again without closing the IDE.

    ✓ Verified by official sourceSource: IDE Release Notes
  • CapabilityQoder

    Qoder shows more informative errors for failed model requests

    Fixed the generic error message shown for some failed model requests. Users now get clearer error information to help identify the source of the problem.

    Usage details

    Who can use it: Qoder 0.5.17; users making model requests.

    How to use it: When a model request fails, open the error details and follow the recommended connection, model, or permission step.

    Tip: Include the error text in your support ticket and address the stated cause instead of retrying blindly.

    ✓ Verified by official sourceSource: Qoder Release Notes
  • CapabilityQoder

    Qoder removed duplicate text from marketplace plugin descriptions

    Fixed a bug that caused promotional text to appear twice in the Description section of some marketplace plugins.

    Usage details

    Who can use it: Qoder 0.5.17; all plans using the marketplace.

    How to use it: Open the plugin details in the marketplace to view the updated Description section.

    Tip: When choosing a plugin, rely on its capabilities and permissions rather than repeated promotional text.

    ✓ Verified by official sourceSource: Qoder Release Notes
  • FeatureQoder

    Qoder added a theme- and language-aware guide for pinned plugins

    The visual guide explaining how to open pinned plugins adapts to the interface language and light or dark theme. It shows how to access pinned plugins from More in the left sidebar.

    Usage details

    Who can use it: Qoder 0.5.17; Qoder desktop users.

    How to use it: Pin a plugin, open More in the left sidebar, and use the pinned plugin entry shown in the guide.

    Tip: If the guide doesn’t appear, upgrade Qoder to the latest version and reopen the theme settings.

    ✓ Verified by official sourceSource: Qoder Release Notes
  • APIQoder CLI

    Qoder CLI adds a protocolVersion option for MCP connections

    A protocolVersion option was added to MCP server configuration, allowing you to specify which MCP protocol version to use for the connection.

    Usage details

    Who can use it: Qoder CLI 1.1.93; MCP server configurations.

    How to use it: Add the `protocolVersion` field to the MCP server definition and specify the protocol version you want; then check the connection with /mcp.

    Tip: Use a version supported by the server; if there’s a compatibility issue, remove the field and try the default negotiation.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • SettingQoder JetBrains Plugin

    Qoder JetBrains Plugin adds three levels of code security scanning

    Code security scanning now has three levels that users can select. The setting has been added under Personal Settings > Security.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; supported JetBrains IDEs.

    How to use it: Open Personal Settings > Security and select one of the three scan levels.

    Tip: Choose the balanced level for an initial scan and the more comprehensive level before a critical release.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • Plugin / IntegrationQoderWake

    QoderWake connects to WeCom channel posts and comments

    You can start a task by @mentioning the bot in a WeCom channel post or comment. Waker uses the readable context of the existing post and sends the result to the same location.

    Usage details

    Who can use it: QoderWake 1.2.22; WeCom channel posts and comments.

    How to use it: Write a post or comment in a WeCom channel, @mention the bot, and specify the task in the message.

    Tip: Keep the task brief and clear; the existing post context is already passed to Waker.

    ✓ Verified by official sourceSource: QoderWake Release Notes
  • CapabilityQoder CLI

    Qoder CLI continues when tool-generated output reaches the limit

    If the model reaches the output token limit while generating tool calls, incomplete calls are dropped and generation resumes automatically.

    Usage details

    Who can use it: Qoder CLI 1.1.93; tool-call turns approaching the output limit.

    How to use it: Update to CLI 1.1.93; automatic continuation at the limit requires no special action.

    Tip: Splitting a task into smaller parts instead of requesting many tools in one turn reduces the need to regenerate.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder CLI

    Qoder CLI fixes an issue saving custom models with thinking support

    Fixed an issue that prevented some custom models with thinking support from being saved because the connection check failed.

    Usage details

    Who can use it: Qoder CLI 1.1.93; custom or BYOK models with thinking support.

    How to use it: Save the model using /model or the relevant custom model configuration; the connection check should no longer block valid configurations.

    Tip: Match the thinking parameters in the custom model to values supported by the provider.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder CLI

    Qoder CLI now cleans up background processes when a Subagent exits

    Fixed a bug that left background processes running when a Subagent finished.

    Usage details

    Who can use it: Qoder CLI 1.1.93; Subagents and background processes.

    How to use it: Update to CLI 1.1.93 or later; processes are automatically cleaned up after a Subagent finishes.

    Tip: To check for processes left behind in older versions, inspect the operating system’s process list at the end of the session.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder CLI

    Qoder CLI reduces tool-call delays during heavy TUI streaming

    Fixed an issue that caused tool calls to be delayed for minutes when the TUI generated a large amount of streaming output.

    Usage details

    Who can use it: Qoder CLI 1.1.93; heavy streaming and tool-call sessions.

    How to use it: Update to CLI 1.1.93; send tool calls as usual during heavy output.

    Tip: Reducing output from particularly noisy commands helps keep the TUI running smoothly.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • CapabilityQoder JetBrains Plugin

    Qoder JetBrains Plugin refreshes context usage during streaming

    The context usage indicator now refreshes in real time while a response is streaming.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; streaming responses.

    How to use it: Watch the context usage indicator while generating a chat response.

    Tip: If the indicator rises quickly, remove unnecessary file attachments and start a new turn.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • FeatureQoder JetBrains Plugin

    Qoder JetBrains Plugin lets you edit and resend previous questions

    Previously submitted questions can now be edited using the edit button at the bottom right of the question card.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; all chat users.

    How to use it: Click the edit button at the bottom right of a previous question card, change the text, and resend it.

    Tip: Editing the same card instead of starting a new chat preserves the current context.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • AvailabilityQoder JetBrains Plugin

    Qoder JetBrains Plugin makes Qoder CLI installation and startup more reliable

    Qoder CLI installation and startup are now more reliable. Canceling installation no longer displays an incorrect failure message.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; JetBrains users using the Qoder CLI integration.

    How to use it: Run the CLI installation from the plugin; if you cancel, don’t treat the operation as a failure unless an actual error message appears.

    Tip: Restarting the editor after installation helps apply PATH and local service changes.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • SettingQoder JetBrains Plugin

    Qoder JetBrains Plugin adds prediction, context, submission, and follow-up settings

    Chat settings now include options for prediction suggestions, context usage display, send shortcut, and handling follow-up messages as queue or steer.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; plugin chats.

    How to use it: Open the plugin’s chat settings; configure prediction suggestions, context usage, send shortcut, and follow-up behavior separately.

    Tip: For long tasks, turning on context usage display and using queue for follow-up messages keeps the workflow uninterrupted.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • AvailabilityQoderWake

    QoderWake fixes installation repair during Windows updates

    Fixed a bug that prevented repair when the Windows installer was rerun during an update through certain installation paths.

    Usage details

    Who can use it: QoderWake 1.2.22; Windows installations.

    How to use it: Update QoderWake to 1.2.22; after updating, you can rerun the installer and use the Repair option.

    Tip: On devices where the installation path has changed, try Repair first, then reinstall.

    ✓ Verified by official sourceSource: QoderWake Release Notes
  • CapabilityQoder CLI

    Qoder CLI fixes the Starting status getting stuck after a Subagent task

    Fixed a bug that caused the progress line to keep showing Starting after a task was delegated to a Subagent.

    Usage details

    Who can use it: Qoder CLI 1.1.93; Subagent tasks.

    How to use it: Update to CLI 1.1.93 and start a Subagent; monitor the actual status in the progress line.

    Tip: If the status doesn’t change, clear the old version’s cache and restart the CLI.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • AvailabilityQoder JetBrains Plugin

    Qoder JetBrains Plugin notifies users when the model changes mid-chat

    Users are notified when the model is changed during an active session.

    Usage details

    Who can use it: JetBrains Plugin 2026.1119.91743910; chats using the model selector.

    How to use it: Choose another model from the model selector during a chat; a notification about the change will appear in the plugin.

    Tip: For long tasks, check the new model’s context and capabilities after switching.

    ✓ Verified by official sourceSource: JetBrains Plugin Release Notes
  • CapabilityQoderWake

    QoderWake links memory and Skill notifications to the correct conversation

    Fixed a bug that caused memory and Skill update notifications to appear in unrelated conversations.

    Usage details

    Who can use it: QoderWake 1.2.22; conversations using memory and Skill.

    How to use it: After updating, open the relevant Waker conversation and check the notifications.

    Tip: When reviewing notifications, check which Waker and conversation ID the message is associated with.

    ✓ Verified by official sourceSource: QoderWake Release Notes
  • CapabilityQoderWake

    QoderWake refreshes the participant list when a WakerFlow is edited

    Fixed a bug that caused participants from previous runs to remain visible after saving edits in the WakerFlow list. The list now reflects the current configuration.

    Usage details

    Who can use it: QoderWake 1.2.22; WakerFlow lists.

    How to use it: Edit and save the WakerFlow, then reopen the list to check that the participants have been updated.

    Tip: Refresh the participant list before starting a new run.

    ✓ Verified by official sourceSource: QoderWake Release Notes

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, 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, weights on Hugging Face, also available on Amazon Bedrock. The old Kimi K3 card on Hugging Face now redirects visitors to K3.1 Pro. Kimi Code 2.7.0 added the `kimi server` command, which runs without opening a browser; 2.7.1, 2.7.2, and Desktop 1.4.1/1.4.2 fixed the selected thinking effort resetting after submission.

Muse Wonder

44
Meta · Meta AI

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

Qwen4 Max

43
Alibaba · Qwen

Released on November 16; 14th with 43 points. Available in Model Studio as `qwen4-max` (1M context, Singapore pricing $1,50 / $4,50), QwenCloud Token Plan, QwenWork, QwenStudio, and OpenRouter (`qwen/qwen4-max`). Available in Qoder from day one. Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

Anthropic’s fast, low-cost model at $0.10 / $0.50. Meta says the new Muse Ember delivers the same score at about 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, available in Meta Model API as `muse-ember-1.0` and on all Muse Code plans.

Sources