YZ·RADARyz.mikail.net
Daily edition · 19 November 2026

Gemini 4 Neon enters third place with 61 points, Muse Wonder Full drops out of the TOP 10

There is just one new entry in the TOP 10 today. Gemini 4 Neon, released by Google on November 18, entered the Artificial Analysis Intelligence Index v5.0.1 directly in third place with 61 points. Claude Fable 5.2 (65) remains at the top, with GPT-6.2 Astra (62) in second place. The scores of the models below Neon are unchanged, but they have all dropped one place: Gemini 4 Carbon (60) is fourth, Claude Opus 5.5 (59) fifth, GPT-6.2 Sol (57) sixth, Claude Sonnet 5.5 (55) seventh, Gemini 4 Hydrogen (54) eighth, and GPT-6.1 Terra (51) ninth. Gemini 4 Argon and Muse Wonder Full are tied at 50 points; Argon takes tenth place because it ranks higher in the Artificial Analysis table, pushing Muse Wonder Full out of the TOP 10. Google now has four models in the TOP 10. Muse Wonder Full (50) now leads the watchlist. It is followed by Kimi K3.1 Pro (46, 12th place, still the strongest open-weight model), Muse Wonder (44), Qwen4 Max (43), Claude Haiku 5.5, and Muse Ember (36). Grok 4.7 (35) has dropped off the watchlist. LMArena’s live page was again blocked by a CAPTCHA; the latest accessible view is dated November 11 and lists Gemini 4 Carbon, Gemini 4 Argon, and Muse Wonder in the top three.

Gemini 4 Neon is generally available in Gemini API, AI Studio, the Gemini app (Pro and Ultra), Antigravity, and Gemini Enterprise. API pricing is $3 / $15, with a 1M-token context window and low, medium, high, and xhigh thinking levels. According to Artificial Analysis, it costs $6,44 per task at the xhigh setting, making it more expensive than Carbon. GitHub Copilot has also made Neon available across all surfaces, including VS Code, JetBrains, Copilot CLI, and coding agent, on the Pro+, Max, Business, and Enterprise plans. Gemini API now accepts Word, PowerPoint, Excel, and OpenDocument files directly; you can select a page, sheet, cell range, time range, or layer to process and pay for only the required portion. Gemini Enterprise’s Canvas is generally available for creating interactive web apps and dashboards without writing code. Tasks in Gemini is now free on Hydrogen and has expanded to the US, India, Canada, Australia, Japan, and South Korea; Türkiye is not on the list. In the US, tasks can now be started from Google Messages and iMessage, or by forwarding an email to gemini@google.com; after Gemini gets approval, it can shop with Google Pay and call businesses on the user’s behalf. AI Studio Build generates full-stack Android apps from a single prompt, including Kotlin, Firestore, and Google Sign-In. Antigravity has added a Robotics extension for Hydrogen. Genie 4 is now available to Ultra subscribers in Project Genie, and the waitlist has opened for Stitch Loop, a continuously running product agent. Custom organization policies for FloorSetting are generally available in Model Armor. The Gemini app is now available on low-memory Android Go devices, personal avatars in Google Vids are free for everyone, and temporary conversations that are not saved to history have been added to AI Mode.

Of Anthropic’s 46 entries, 44 are from Claude Code v2.1.328. Auto mode for Opus 4.6 and Sonnet 4.6 now also works on Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry, and Claude apps gateway; admins can disable it with `disableAutoMode`. On the security front, if the sandbox cannot start, the next sandboxed Bash call is rejected rather than silently running without protection. The built-in `gh api` is now disabled by default in self-hosted runners. The `mcp_tool` hooks now support `onFailure: "block"`, and `deniedMcpServers` rules are also enforced in cloud and self-hosted sessions. `/code-review` now focuses primarily on reporting bugs and CLAUDE.md violations, labeling findings as Important, Pre-existing, and Nit; `/simplify` is recommended for simplification. In Code Mode, the 0.75-second wait after scripts has been removed, and resuming large sessions is faster. On the OpenAI side, ChatGPT’s Live Dictation feature streams text as you speak, and the 10-minute recording limit has been removed; it is rolling out across all plans on iOS, Android, web, and desktop. Regional data and inference residency can now be selected on a self-serve basis for API projects; per-request region selection can also be enabled for Global projects. A content partnership with Le Groupe Figaro was announced. openai-node 7.60.0 added MCP server configuration to Live API sessions, while removing Managed Agents budget controls and events from the SDK surface; TypeScript code using these may no longer compile. Agents SDK Python 0.25.0 no longer loses token usage from failed Responses calls. At Meta, Loblaw became Muse’s first retail connector in Canada, and Muse Voice launched on mobile, web, and Mac in Canada.

xAI’s 20 entries are from Grok Bot’s November 18 updates. Connector tools now have four permission levels: Always allow, Auto-review, Needs approval, and Blocked. The built-in GitHub connector is now subject to the team MCP allowlist; it stays disabled if no rule is set. The team data retention period now also deletes meeting transcripts and notes. Recording is limited to people aged 18 and older. One change to watch: for users with a saved card who have not previously set a rule, the Buy credits screen now loads with automatic top-up enabled. On Alibaba’s side, Qoder added day-one support for Qwen4-Max and lowered credit multipliers starting November 18: Auto is now 0,3x, Fast 0,6x, and Efficient 0,1x, while Ultimate remains at 2,0x. Qoder also lowered the default thinking effort for GLM-5.3, Kimi-K3.1-Pro, Kimi-K2.8-Preview, and DeepSeek-Flash from Maximum to High. QoderWake’s digital employees can respond to eligible group messages without an @ mention. With Qoder CLI 1.1.91, enterprise admins can distribute Skill and MCP configurations to members. DeepSeek has not charged extra for server-side web searches and page reads triggered within a model call since November 18 at 12:00 (Beijing time); the independent Search and Fetch APIs remain priced at $0,75 per 1.000 successful requests. Harness v0.2.2-rc.5 routes DeepSeek models to this free search. Tencent released the Hy4 flash, Hy4 flash-FP8, and Hy4 flash-Base weights on Hugging Face, ModelScope, and GitCode. LoopForge added writable workspaces for personal conversations, Antigravity account sign-in, host-based Git tokens, and authenticated HTTP MCP support; LoopForge Local was marked legacy. AI-Infra-Guard v4.6.11 introduced the CoTInjection attack operator. Moonshot’s Kimi Work 3.2.28 added project-level AGENTS.md and default PPTX delivery; the enterprise SSO news has not been verified by an official source. Xiaomi MiMo Code split media reading into the `read_image`, `read_audio`, and `read_video` tools; the change is not yet in a tagged release. Z.ai had no new-user changes between November 17 and 19. A total of 165 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, it leads the ranking with 65 points. Available on all paid Claude plans; `claude-fable-5-2` in the API, $8 / $40, with thinking always adaptive. The Claude Meetings beta opened to Team and Enterprise plans on November 12, initially for Google Meet.Released: 9 November 2026

  2. 02

    GPT-6.2 Astra

    OpenAI · ChatGPT=
    62pts

    Released on November 17, it holds second place with 62 points. The cost per task is $4,00 at the Max setting. Available in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans; `gpt-6.2-astra` in the API, $8 / $40, with cached input at $0,40. 1,05M context, 128K output; reasoning effort `none` is not accepted. Also available on GitHub Copilot (Pro+, Max, Business, Enterprise).Released: 17 November 2026

  3. 03

    Gemini 4 Neon

    Google · GeminiNEW
    61pts

    Released on November 18, it entered directly in third place with 61 points. The cost per task is $6,44 at the xhigh setting. `gemini-4-neon` in the API (paid tier only), $3 / $15, 1M context; offers low, medium, high, and xhigh thinking levels. Available in the Gemini app on Pro and Ultra, on Antigravity’s paid Pro/Ultra plans, and in Gemini Enterprise when enabled by an admin. Generally available on GitHub Copilot for Pro+, Max, Business, and Enterprise.Released: 18 November 2026

  4. 04

    Gemini 4 Carbon

    Google · Gemini▼ 1
    60pts

    Its score is unchanged (60), but it dropped to fourth place with the arrival of Gemini 4 Neon. `gemini-4-carbon` in Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, $4 / $20 starting December 1). Carbon works with CodeMender to run the security loop in Google’s new Stitch Loop agent.Released: 28 October 2026

  5. 05

    Claude Opus 5.5

    Anthropic · Claude▼ 1
    59pts

    Its score is unchanged (59), but it dropped to fifth place with the arrival of Gemini 4 Neon. 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▼ 1
    57pts

    Released on November 17, its score is unchanged (57), but it dropped to sixth place with the arrival of Gemini 4 Neon. The cost per task is $1,01 at the Max setting, one-quarter of Astra’s. `gpt-6.2-sol` in the API, $2 / $10, with cached input at $0,10; 1,05M context, 128K output. Available in ChatGPT Work, Codex, and GitHub Copilot.Released: 17 November 2026

  7. 07

    Claude Sonnet 5.5

    Anthropic · Claude▼ 1
    55pts

    The strongest model also available on claude.ai’s free plan; its score is unchanged (55), but it dropped to seventh place.Released: 28 September 2026

  8. 08

    Gemini 4 Hydrogen

    Google · Gemini▼ 1
    54pts

    Released on November 12, it is eighth with 54 points (down one place, score unchanged). Costs $3,16 per task at the xhigh setting. `gemini-4-hydrogen` is also available on the free API tier, with paid pricing at $1,50 / $7,50. Tasks in Gemini has been available on Hydrogen and free in six countries since November 17; Antigravity added a Robotics extension for Hydrogen.Released: 12 November 2026

  9. 09

    GPT-6.1 Terra

    OpenAI · ChatGPT▼ 1
    51pts

    Released on October 26 in ChatGPT Work, Codex, and the API; focused on vision and computer use. Its score is unchanged (51), but it dropped to ninth place. There is no GPT-6.2 version yet.Released: 26 October 2026

  10. 10

    Gemini 4 Argon

    Google · Gemini▼ 1
    50pts

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

Key signals

  • Gemini 4 Neon enters third place with 61 points; Muse Wonder Full (50) drops out of the TOP 10.
  • The seven models below Neon each drop one place without a change in score; Claude Fable 5.2 (65) remains at the top.
  • Claude Code 2.1.328: Auto mode for Opus/Sonnet 4.6 in the cloud; `gh api` disabled by default in self-hosted runners.
  • DeepSeek: server-side web search within a model call is no longer charged separately.
  • Qoder lowered credit multipliers, Tencent Hy4 flash became open-weight, and automatic top-up is enabled by default in Grok Bot.
  • 165 new developments: Anthropic 46, Alibaba 23, Google 23, Tencent 23, xAI 20, OpenAI 11, Moonshot 10, DeepSeek 4, Xiaomi 3, Meta 2.

Worth trying today

01

Use Gemini 4 Neon with the right thinking level

Set the model field to `gemini-4-neon` in Gemini API; the model is available only on the paid tier and costs $3 / $15. Choose the thinking level to match the task: low or medium for short questions, high for multi-step coding and analysis, and xhigh for the hardest tasks. Artificial Analysis’ score of 61 is for the xhigh setting, which costs $6,44 per task, so don’t use xhigh for every request. You can now send Word, Excel, PowerPoint, or OpenDocument files directly; use page, sheet, or cell-range slicing to process and pay for only the required portion. If you use GitHub Copilot (Pro+, Max, Business, Enterprise), select Gemini 4 Neon from the model picker. In Business and Enterprise, an admin must first enable the model under Copilot settings > model policy.

Google · Gemini
02

Review Claude Code 2.1.328’s security settings

Run `claude update` to upgrade to 2.1.328. If you use a self-hosted runner, the built-in `gh api` is now disabled by default; if GitHub API access is genuinely needed, start the runner with `--use-anthropic-gh true`. If the sandbox cannot start, the next sandboxed Bash call is now rejected; if you want the session not to start at all in this case, add `sandbox.failIfUnavailable` to the settings. To stop execution when an MCP tool hook fails, add `"onFailure": "block"` to the relevant `mcp_tool` hook. To block a claude.ai connector everywhere, add the `claude.ai <ad>` entry to the `deniedMcpServers` list; this rule now also applies to cloud and self-hosted sessions. Run `/code-review` for code review and `/simplify` for simplification.

Anthropic · Claude
03

Run web searches through DeepSeek model calls

Since November 18 at 12:00 (Beijing time), server-side search and page reading triggered during a model call are no longer charged separately; you pay only for tokens. Add `deepseek_web_search` or `deepseek_read_url` to `tools` in Chat Completions, Messages, or Anthropic-compatible requests. In the Responses API, use `tools: [{"type": "web_search"}]` with a supported model. The standalone `/v1/tools/search` and `/v1/tools/fetch` endpoints remain priced at $0,75 per 1.000 successful requests; if your code calls these endpoints, consider moving the search to the tool within the model. If you use DeepSeek Harness, upgrade to v0.2.2-rc.5; DeepSeek models automatically route search to the free server-side option.

DeepSeek · DeepSeek
04

Update your model selection for Qoder’s new credit multipliers

Since November 18, Qoder has used multipliers of 0,3x for Auto, 0,6x for Fast, and 0,1x for Efficient; Ultimate remains at 2,0x, and your current credit balance is unchanged. Choose Efficient for simple edits, explanations, and small tests; credit consumption is now one-third of what it was. Qoder also lowered the default thinking effort for GLM-5.3, Kimi-K3.1-Pro, Kimi-K2.8-Preview, and DeepSeek-Flash from Maximum to High. If you use these models for difficult tasks, open Manage models from the model selector below the input box and set the effort back to Maximum. Qwen4-Max is also available in the model picker from day one; update Qoder to select it.

Alibaba · Qwen

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 · 46

  • AvailabilityClaude Code

    Auto mode comes to cloud providers for Opus 4.6 and Sonnet 4.6

    Auto mode is now available with Claude Opus 4.6 and Claude Sonnet 4.6 on Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry, and Claude apps gateway. Interactive sessions started without a specified permission mode can start in Auto mode with these models; prompts on third-party providers also offer Auto mode for plan and Bash prompts.

    Usage details

    Who can use it: Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry, and Claude apps gateway; Claude Code sessions using the relevant models.

    How to use it: Start Claude Code with the relevant model on a supported provider without specifying a permission mode; select Auto mode in prompts. Use `disableAutoMode` to turn it off through an administrator setting.

    Tip: Test how provider and organization policies are applied in a small repository before using automatic permission decisions in production workflows.

    ✓ Verified by official sourceModels: Claude Opus 4.6, Claude Sonnet 4.6Source: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code

    The next Bash call is explicitly rejected if the sandbox cannot start

    If the sandbox cannot start and `sandbox.failIfUnavailable` is not enabled, the next sandboxed Bash call is now rejected with a message stating that the sandbox is disabled for the session. This makes the situation visible instead of silently reducing protections.

    Usage details

    Who can use it: Claude Code sessions using the Bash sandbox.

    How to use it: Make a Bash call in an environment where the sandbox cannot start; check the warning and the session’s sandbox status. Use `sandbox.failIfUnavailable` to fail explicitly.

    Tip: If your production security policy does not allow operation without a sandbox, enable `sandbox.failIfUnavailable`.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Plugin / IntegrationClaude Code Self-hosted Runner

    Self-hosted runner gh api access is disabled by default

    In self-hosted runner environments where settings protect Claude GitHub tools by name, the built-in `gh api` is now disabled by default. The `--use-anthropic-gh true` option is required to enable it.

    Usage details

    Who can use it: Self-hosted runners and organizations that protect GitHub tools with settings.

    How to use it: After confirming it is required, add `--use-anthropic-gh true` to the runner startup options.

    Tip: Assess whether to enable it in conjunction with the organization’s GitHub permissions policy; treat the disabled-by-default behavior as a secure baseline.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Pricing / PlanClaude Code

    Plugin usage totals no longer stop at 2,000 tokens on 1M context models

    The plugin total in `/usage` no longer stops at 2,000 tokens on 1M context models. Low token estimates in `/usage` and `/plugin` when no model is selected have also been corrected.

    Usage details

    Who can use it: Claude Code `/usage` and `/plugin` screens; includes 1M context models.

    How to use it: Run a plugin with a 1M context model and check the totals in the `/usage` and `/plugin` screens.

    Tip: When comparing plugin costs, don't rely only on the initial 2,000-token indicator; use the current total.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AgentClaude Code

    Subagent progress summaries are generated with a smaller, faster model

    In terminal sessions using the Anthropic API, subagent progress summaries are now generated from brief notes using the default small, fast model. This change also uses fewer tokens when telemetry is disabled.

    Usage details

    Who can use it: Claude Code terminal sessions and subagents; Anthropic API.

    How to use it: Run a subagent and monitor its progress summaries; compare summary token usage in the usage screen.

    Tip: Keep this default in workflows that use many subagents and reserve the main budget for the task model.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AgentClaude Code Agent SDK

    Subagents now wait for their own subagents by default

    When `CLAUDE_CODE_FORK_SUBAGENT=0` is set, subagents in `claude -p`, Agent SDK, Desktop, and VS Code now wait for their own subagents to complete by default. This reduces the risk of the parent session ending too early.

    Usage details

    Who can use it: Claude Code `-p`, Agent SDK, Claude Desktop, and VS Code; workflows using `CLAUDE_CODE_FORK_SUBAGENT=0`.

    How to use it: Set `CLAUDE_CODE_FORK_SUBAGENT=0` in the environment and run a subagent chain; observe the wait behavior in the session result.

    Tip: Plan for the total wait time separately when using long subagent chains.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code Artifacts

    Artifact access status is shown more accurately in shared folders

    Claude no longer mistakenly reports an artifact read from a shared folder or published by another organization as private. Its answer about who can open the artifact also takes folder access into account.

    Usage details

    Who can use it: Claude Code Artifact workflows and shared folders.

    How to use it: Read an artifact in a shared folder or publish one to another organization; check the access view and Claude's explanation.

    Tip: Before deciding whether to share, verify the artifact folder's actual membership settings rather than relying on Claude's text.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code

    Large session resume speed and memory use improved

    Resume time for large sessions has been reduced, and memory use in sessions containing many images has been lowered. Responsiveness has also improved in long SDK and Desktop sessions running slow tools or commands.

    Usage details

    Who can use it: Claude Code, Claude Desktop, and Agent SDK sessions.

    How to use it: Resume a large session or one containing images, then compare startup time and memory use with the previous version.

    Tip: Continue to archive very large transcripts regularly; this improvement does not guarantee unlimited memory.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code CLI

    claude -p counts background subagent wait time as a single 10-minute window

    `claude -p` running with `CLAUDE_CODE_FORK_SUBAGENT=0` includes waiting for background subagents and the subagents they are waiting on in a single 10-minute wait period. `CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS` can be increased for longer chains.

    Usage details

    Who can use it: Headless `claude -p` and non-forking subagent chains.

    How to use it: If needed, increase `CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS` and run `claude -p`.

    Tip: Set the value according to the longest expected subagent chain rather than increasing it without limit.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • RetiredClaude in Chrome

    Claude in Chrome shortcut tools removed from Claude Code

    The built-in Claude in Chrome tools `shortcuts_list` and `shortcuts_execute` have been removed from Claude Code. These tools listed and ran shortcuts saved in the extension.

    Usage details

    Who can use it: Claude Code's built-in Claude in Chrome tools.

    How to use it: Remove automations that call these two tools or migrate them to the current user workflow supported by the extension.

    Tip: Before updating, search your commands, skills, and integrations for these tool names.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Plugin / IntegrationClaude in Chrome

    Claude in Chrome makes it clearer that it asks permission before acting

    The Claude in Chrome welcome screen and `/chrome` now explain that Claude asks before taking action on a site. This prompt can be skipped if the Permission mode or a `/permissions` rule has already determined the decision.

    Usage details

    Who can use it: Claude in Chrome workflow in Claude Code.

    How to use it: Open the `/chrome` command and review the behavior before actions; manage automatic decisions through `/permissions`.

    Tip: Use a mode that explicitly requests permission on untrusted sites.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AvailabilityClaude Code Cloud

    Cloud sessions preserve summaries and recent messages saved late

    When a message is saved late, cloud sessions no longer lose the compaction summary or recent messages upon resuming. This reduces the risk of context loss in long cloud sessions.

    Usage details

    Who can use it: Claude Code cloud sessions.

    How to use it: Resume a long cloud session and check that recent messages and the compaction summary are preserved.

    Tip: For critical work, separately verify the last response and changed files before resuming the session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code

    Warnings indicate when messages may be held or dropped in cloud sessions

    Users are now notified that a message sent from a session that skips permission prompts may be held or dropped by a cloud session. Claude Code also explains what to do about this in `claude -p` output.

    Usage details

    Who can use it: Claude Code cloud sessions and cross-session messaging; sessions that skip permission prompts.

    How to use it: Send a message to a cloud session from another session and check the terminal warning or `claude -p` stderr output.

    Tip: To avoid losing critical messages, use the permission path in the session's settings after a warning says the message was held.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • FeatureClaude Code Code Mode

    Code Mode is faster and more observable at runtime

    The 0.75-second delay after most scripts has been removed in Code Mode; MCP server connection status and changes to the server list are passed to Claude. The commands run by scripts and their output are now visible, and editing and development server workflows have also been improved.

    Usage details

    Who can use it: Claude Code Code Mode; including auto and bypass-permissions workflows.

    How to use it: Enable Code Mode, run a script in a session with an MCP server, and monitor the command/output flow.

    Tip: Start development servers as background tasks that can be monitored under `/tasks` in Code Mode.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • FeatureClaude Code Code Mode

    Code Mode script output no longer triggers UserPromptSubmit hooks

    Code Mode no longer runs UserPromptSubmit hooks on script output, including managed hooks. A PostToolUse hook that uses `updatedToolOutput` is recommended for masking or modifying output.

    Usage details

    Who can use it: Claude Code Code Mode; user and managed hooks.

    How to use it: If script output needs to be modified, define a PostToolUse hook and use the `updatedToolOutput` field.

    Tip: Move sensitive-data cleanup to the PostToolUse stage instead of UserPromptSubmit.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • FeatureClaude Code

    Code review findings narrowed and severity labels added

    For most models, `/code-review` now primarily reports bugs and CLAUDE.md violations; `/simplify` is recommended for reuse and simplification. Findings are labeled Important, Pre-existing, or Nit, and the recommended fix appears at the end of the summary.

    Usage details

    Who can use it: Claude Code `/code-review`; most supported models.

    How to use it: Run `/code-review` for reviews, and run `/simplify` separately for reuse or simplification cleanup.

    Tip: In CI reviews, use only Important findings as a gate and track Pre-existing and Nit results separately.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code computer use

    Computer use approval dialog shows access levels more clearly

    The computer use approval dialog now shows each app's access level and clearer warnings. Users can more easily see the scope of access being granted before approving.

    Usage details

    Who can use it: Claude Code computer use approval flow.

    How to use it: When the computer use approval dialog opens, review each app's access level and choose whether to approve or deny it.

    Tip: Approve only the access level required rather than granting broad access to apps that do not need it.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code

    Missing OpenTelemetry endpoints are reported in status and doctor

    If the OpenTelemetry exporter has no endpoint and data is likely to be lost, `/status` and `claude doctor` now display a warning. This lets you check from the command line whether telemetry is actually being sent.

    Usage details

    Who can use it: Claude Code installations using OpenTelemetry.

    How to use it: Run `/status` or `claude doctor`; if you see a warning, complete the telemetry exporter's endpoint configuration.

    Tip: If you expect observability in production, include this warning in your CI or environment health check.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code CLI

    A variable to disable deferred tool calls from running on resume has been added

    `CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME` was added to prevent host applications from executing a tool call deferred by a hook during `claude -p --resume` and `--continue`.

    Usage details

    Who can use it: Claude Code CLI, especially headless sessions run by a host, SDK, or automation.

    How to use it: Set `CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME=1` in the host environment, then run `claude -p --resume` or `--continue`.

    Tip: Make this variable the default security option if hook decisions must always be reevaluated when resuming a session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Plugin / IntegrationClaude Code

    Embedded applications can now restrict managed mods through policy alone

    Applications embedding Claude Code can enable the policy plugin's `allowManagedModsOnly` option. Admin-enforced managed settings retain precedence as specified by this option.

    Usage details

    Who can use it: Applications embedding Claude Code and organization policies.

    How to use it: Enable `allowManagedModsOnly` in the embedded application's policy plugin; also review the organization's managed settings.

    Tip: When disabling user-installed mods, document the allowlisted managed mods separately.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Custom repository folder name added for capacity 1 runner sessions

    `--repository-folder` was added to change the name of the repository path's base folder when using `--capacity 1` on a self-hosted runner. This lets you use the selected folder name instead of the default path layout, which cannot be a symbolic link.

    Usage details

    Who can use it: Claude Code self-hosted runner; repositories running with `--capacity 1`.

    How to use it: Add `--capacity 1 --repository-folder <folder-name>` to the runner startup configuration.

    Tip: Use consistent, meaningful folder names across multiple runner environments to make checkout paths easier to track.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Plugin / IntegrationClaude Code MCP

    Complex MCP schemas no longer disappear with third-party providers

    MCP tools whose input schemas begin with `anyOf`, `oneOf`, or `allOf` no longer become invisible with third-party providers or when telemetry/feature flags are disabled. If a schema is invalid, the tool is omitted and Claude is given the reason.

    Usage details

    Who can use it: Third-party providers and MCP sessions with telemetry/feature flags disabled.

    How to use it: Publish the MCP tool's input schema with complex unions; check its visibility in `/mcp` or the tool list.

    Tip: Monitor omission reasons in MCP server logs so invalid schemas don't prevent the call entirely.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Pricing / PlanClaude Code

    Running out of usage credits is no longer labeled a rate limit

    The error shown when usage credits run out during a response has been corrected; the message no longer labels the issue as a rate limit. This distinguishes exhausted credits from a rate limit.

    Usage details

    Who can use it: Claude Code plans and sessions that use credits.

    How to use it: If credits run out mid-response, check the error message and the `/usage` view.

    Tip: If you run out of credits, check your plan or spending options before retrying.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • Plugin / IntegrationClaude Code MCP

    MCP denial rules now apply to cloud and self-hosted sessions

    Entries in `deniedMcpServers` in the format `claude.ai <name>` now block the relevant connector in cloud and self-hosted runner sessions as well. Previously, this scope did not apply to some session types.

    Usage details

    Who can use it: Cloud and self-hosted runner sessions; `claude.ai` connectors.

    How to use it: Add a `claude.ai <name>` entry for the connector you want to block to the `deniedMcpServers` list in your settings file.

    Tip: Standardize connector naming across your organization; verify rule matching with `/status` and a test session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code

    MCP tool hooks can block an operation when a call fails

    The `onFailure: "block"` option has been added for `mcp_tool` hooks. If a hook fails, the relevant operation is blocked, as with command and HTTP hooks.

    Usage details

    Who can use it: Claude Code hook configurations; MCP tool calls.

    How to use it: Add `"onFailure": "block"` to the relevant `mcp_tool` hook in your settings or hook configuration.

    Tip: Use this option for hooks that perform security or compliance checks; account for temporary network errors in the hook stopping the operation.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code

    Commands rewritten to run outside the sandbox via PermissionRequest run outside the sandbox

    If a PermissionRequest hook rewrites a shell command to run outside the sandbox, Claude Code now executes that command outside the sandbox. This makes the hook’s return value consistent with the actual execution environment.

    Usage details

    Who can use it: PermissionRequest hooks and Bash commands.

    How to use it: Use the outside-sandbox command rewrite behavior in your PermissionRequest hook; verify the command’s execution environment in a test.

    Tip: Use this behavior only with trusted hooks, and log outside-sandbox permissions.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code

    Managed PostToolUse does not run for rejected tool calls

    Calls rejected before execution no longer trigger managed PostToolUse hooks, so they produce no log, rewrite, block, or stop result. In most cases, call information is left to PostToolBatch.

    Usage details

    Who can use it: Managed PostToolUse and PostToolBatch hooks.

    How to use it: When designing auditing for rejected calls, handle PostToolBatch records as well as PostToolUse records.

    Tip: Define a separate path for rejected calls in compliance logs; do not record a tool that never ran as having run successfully.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Claude processes launched by wrappers outside the runner no longer use managed git

    On a self-hosted runner, `claude` launched by a wrapper before or after the session's own Claude process now runs without Anthropic-managed git. Managed git is available only during the session's active lifecycle.

    Usage details

    Who can use it: Self-hosted runner wrapper and hook environments.

    How to use it: Manage git access separately for Claude commands run before and after the session in your wrapper script.

    Tip: Explicitly manage the provider credentials required for out-of-session git operations under a separate security policy.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Desktop Agent SDK

    Desktop and SDK turns that end with an open question no longer continue

    For third-party providers or environments with telemetry/feature flags disabled, Claude Desktop and Agent SDK sessions closed while a question is open now end the turn instead of continuing without a user response.

    Usage details

    Who can use it: Claude Desktop and Agent SDK sessions; third-party providers or environments with telemetry/feature flags disabled.

    How to use it: Close the session while a question is open and verify that the question has been terminated before starting a new turn.

    Tip: In automations, detect pending questions and add logic to restart the session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code

    Spinner and subagent token counts corrected during retries

    Spinner and subagent token indicators no longer count retried responses twice, miss non-streaming responses, or freeze while Claude is thinking. This makes the usage view better reflect actual consumption.

    Usage details

    Who can use it: Claude Code usage and subagent indicators.

    How to use it: Monitor the spinner and subagent usage indicators during a response; compare the total for a session with retries against `/usage`.

    Tip: For cost tracking, rely on completed session usage rather than the live spinner.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code Self-hosted Runner

    Resumed self-hosted sessions check whether previous work is missing

    When a self-hosted runner session is resumed, Claude now checks whether the previous work is still available and reports anything that's missing. This avoids silently assuming that checkouts or files are still present when they've been lost.

    Usage details

    Who can use it: Sessions resumed with a self-hosted runner.

    How to use it: Reopen a self-hosted session and review Claude's previous-work check and missing-file notification.

    Tip: Also verify important files and commits after runner cleanup or redeployment.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Desktop

    Large tool output no longer permanently ends sessions with third-party providers

    A fix addresses some Claude Desktop sessions with third-party providers permanently ending with a `Prompt is too long` error after large tool results. The largest results are now omitted so the session can continue.

    Usage details

    Who can use it: Claude Desktop; third-party provider sessions outside Claude apps.

    How to use it: Test a Desktop session that produces large tool output with a third-party provider; check how the session continues when the context limit is exceeded.

    Tip: Still, paginate tool output and avoid including unnecessarily large results in prompts.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code Scheduled Tasks

    Saving scheduled tasks no longer deletes other tasks

    If the task file cannot be read when using third-party providers or when telemetry/feature flags are disabled, saving no longer deletes the project’s other tasks. A task with a typo in the file is also no longer silently deleted on the next save.

    Usage details

    Who can use it: Claude Code project scheduled tasks; third-party providers or limited-telemetry environments.

    How to use it: Save project scheduled tasks and check that existing tasks are preserved when the file cannot be read.

    Tip: Put the task file under version control; verify the save result until you fix the typo warning.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code

    A Yes/No confirmation appears before submitting a bug report

    The `/bug` command now asks Yes or No before submitting a report. This prevents repeated Enter presses from accidentally submitting a report or opening GitHub.

    Usage details

    Who can use it: Claude Code terminal; `/bug` command.

    How to use it: Run `/bug` and explicitly select Yes when asked whether to submit the report.

    Tip: Check whether the report contains sensitive data before confirming.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code apps gateway

    Invalid maxEffortLevel values now trigger a startup warning

    The Claude apps gateway now displays a startup warning if `maxEffortLevel` in its policy is not a valid effort level. This makes it easier to catch when a policy value is silently ignored.

    Usage details

    Who can use it: Organization policies using the Claude apps gateway.

    How to use it: Check the gateway startup logs and update the policy's `maxEffortLevel` to a valid effort level.

    Tip: Validate effort values against a schema before deploying policies.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AgentClaude Code

    The list-agents list is now shown only to the user

    `/list-agents` now shows the subagent list only to the user; Claude lists sessions itself when needed. This prevents the agent list from being unnecessarily added to model context.

    Usage details

    Who can use it: Claude Code terminal; `/list-agents`.

    How to use it: Run `/list-agents` and check that the list remains in the user interface.

    Tip: If Claude needs information about agent status, ask it explicitly about the relevant session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AvailabilityClaude Code Remote Control

    The session list explains when Remote Control is off

    If a cloud session or another computer is not visible, Claude now explains more clearly that Remote Control is off and why it is not listed. This distinguishes a lack of connectivity from a lack of sessions.

    Usage details

    Who can use it: Claude Code cloud sessions and Remote Control.

    How to use it: When access to a session on the cloud or another computer is not listed, check the Remote Control explanation shown.

    Tip: In environments where connectivity is expected, verify Remote Control status as a health check before starting a session.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Two session UUID variables are now available together in runner hooks

    `CLAUDE_RUNNER_SESSION_UUID` and `CLAUDE_CODE_REMOTE_SESSION_UUID` are now both set for the session itself and in checkout, post-session, and spawn-runner hooks.

    Usage details

    Who can use it: Self-hosted runner hooks.

    How to use it: Read both environment variables in your hook code, and verify that their values refer to the same session before removing the old single-variable assumption.

    Tip: Use one UUID as the primary key in traceability records and retain the other for backward compatibility.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Two metrics added for runner queue and abandoned sessions

    The orchestrator now measures queue wait time by reason with `claude_code_self_hosted_orchestrator_session_queue_wait_by_reason_seconds` and sessions Anthropic chose not to start with `claude_code_self_hosted_orchestrator_sessions_given_up_total`.

    Usage details

    Who can use it: The self-hosted runner orchestrator `/metrics` endpoint.

    How to use it: Collect the orchestrator's `/metrics` output and monitor both metrics by reason, status, and time series.

    Tip: Add the new metrics to your existing runner success-rate and queue-duration dashboards.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Metadata can be added to a self-hosted session for spawn-runner

    The `--metadata key=value` option has been added to `claude --environment`. This metadata helps the `spawn-runner` hook label the session.

    Usage details

    Who can use it: Claude Code self-hosted runner environments.

    How to use it: Start the session using `claude --environment --metadata key=value`.

    Tip: Do not put secrets in metadata; use this field only for routing and observability labels.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • APIClaude Code Self-hosted Runner

    Selected session label can be passed to self-hosted runner hooks

    The `--hook-label` option was added to the orchestrator. The specified label is passed to the `spawn-runner` hook and can be used in runner startup logic.

    Usage details

    Who can use it: Self-hosted runner orchestrator and `spawn-runner` hooks.

    How to use it: Add `--hook-label <label>` to the `claude self-hosted-runner orchestrator` command.

    Tip: Standardize labels using a limited vocabulary, such as team, environment, or priority.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • AvailabilityClaude Code VS Code extension

    VS Code shows a settings link for held-session messages

    When a message from another session is held, the VS Code extension now shows a notification and a link to the settings file where this can be enabled.

    Usage details

    Who can use it: Claude Code VS Code extension and cross-session messaging.

    How to use it: Send a message from another session; open the settings file link in the notification and review the option in line with your organization's policy.

    Tip: Test hold and drop behavior in cloud sessions before changing your organization's settings.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • CapabilityClaude Code

    Pasted text is now clearly marked for Claude

    When telemetry or feature flags are disabled, or when using a third-party provider, pasted text is marked as pasted for Claude. This improves the distinction between user-written and pasted content.

    Usage details

    Who can use it: Claude Code; third-party providers and sessions with telemetry/feature flags disabled.

    How to use it: Paste text into the terminal and submit the prompt; verify in Claude’s context that it is handled as pasted content.

    Tip: Before pasting long text or text from an external source, remove any sensitive data.

    ✓ Verified by official sourceSource: Claude Code v2.1.328 sürüm notları
  • SettingClaude Code CLI

    claude purge no longer hides deletion prompts when output is redirected

    Deletion confirmations remain visible in the terminal when `claude purge` stdout is redirected. This prevents an unseen Enter keypress from deleting project data.

    Usage details

    Who can use it: Claude Code CLI; `claude purge` runs with redirected stdout.

    How to use it: Run `claude purge` with stdout redirected and verify that deletion prompts are clearly visible in the terminal.

    Tip: Before automating deletions, test the interactive confirmation behavior in a separate test environment.

    ✓ Verified by official sourceSource: Claude Code v2.1.327 sürüm notları
  • CapabilityClaude Code

    Unloaded earlier messages are now reported when resuming a session

    If some earlier messages can't be loaded when a session is resumed, Claude Code now reports this explicitly. This prevents missing context from appearing as a silent change in behavior.

    Usage details

    Who can use it: Claude Code 2.1.327 and later; terminal sessions.

    How to use it: Resume a session with `--resume`; check the warning shown if earlier messages can't be loaded.

    Tip: If you see a warning about missing messages, add important context to the new prompt again.

    ✓ Verified by official sourceSource: Claude Code v2.1.327 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 · 11

  • APIOpenAI API Platform

    Self-serve regional data residency selection now available for API projects

    Eligible API organization owners can choose a fixed regional data and inference residency when creating a new project. In a new Global project, they can enable per-request data residency to select the region for each request using a regional API domain.

    Usage details

    Who can use it: Eligible API organizations and supported regions; eligible models and endpoints vary by region, and OpenAI determines access eligibility.

    How to use it: In API Platform, open Projects > Create; select a region under Data and inference residency. To select a region per request, leave Global selected and enable Enable per-request data residency, then use the API key with the relevant regional API domain.

    Tip: The residency region cannot be changed after project creation; check the region and supported models/endpoints in advance.

  • CapabilityChatGPT

    Live Dictation in ChatGPT now streams text live and removes the 10-minute limit

    Live Dictation converts speech to text as you speak, reducing the wait for dictation to finish. OpenAI has also removed the 10-minute recording limit.

    Usage details

    Who can use it: Rolling out globally to all ChatGPT plans on iOS, Android, web, and desktop apps.

    How to use it: Tap or click the dictation/microphone button in the message composer; tap or click the waveform to show or hide the streaming text.

    Tip: Before starting a long recording, turn on live text and check critical names and numbers as you speak.

  • Plugin / IntegrationChatGPT

    Discover Le Figaro journalism in ChatGPT through summaries, quotes, and links

    The partnership between OpenAI and Le Groupe Figaro makes current news and archival content from Le Figaro and its affiliated publications discoverable in ChatGPT through summaries and quotes. Results will include links to the original articles; Le Figaro also plans to use ChatGPT Enterprise across its teams.

    Usage details

    Who can use it: News, business, culture, and lifestyle content from Le Figaro and its affiliated publications; rollout scope and phased deployment details have not been announced.

    How to use it: Search in ChatGPT for Le Figaro or a related topic; when content appears, read the summary and open the link to the original article.

    Tip: Do not rely on the summary alone to make decisions; open the linked original article to verify the context and publication date.

  • APIOpenAI Node SDK

    Node SDK adds Live MCP server configuration

    openai-node v7.60.0 adds API support for MCP server configuration in Live API sessions. Node applications can now set up the Live stream with MCP server settings.

    Usage details

    Who can use it: openai-node v7.60.0 and later; Node applications using the Live API.

    How to use it: Upgrade openai-node to v7.60.0 or later and use the new Live MCP server configuration when setting up a Live API session.

    Tip: Test authentication, tool scope, and disconnection scenarios before deploying the MCP server connection to production.

    ✓ Verified by official sourceSource: openai-node v7.60.0 sürüm notları
  • RetiredOpenAI Node SDK

    Node SDK hides Managed Agents budget controls and events

    openai-node v7.60.0 hides Managed Agents session budget controls and related events from the SDK surface. This is an API surface change that requires a compatibility check for Node applications using the budget events added in the November 17 typings.

    Usage details

    Who can use it: openai-node v7.60.0 and later; Node applications using Managed Agents.

    How to use it: Run the TypeScript build before upgrading to v7.60.0; remove imports for budget controls/events that are no longer exported, or adapt them to the current Managed Agents API surface.

    Tip: Pin the SDK version and run a separate build test for Managed Agents types with every upgrade.

    ✓ Verified by official sourceSource: openai-node v7.60.0 sürüm notları
  • APIOpenAI Node SDK

    Node SDK adds standalone CSAM moderation types

    openai-node v7.60.0 adds type definitions for the standalone CSAM moderation flow. This is new developer support in the Node SDK for the OpenAI CSAM Image Classifier API launched on November 16.

    Usage details

    Who can use it: openai-node v7.60.0 and later; use of the CSAM Image Classifier is limited to approved organizations.

    How to use it: Upgrade the SDK to v7.60.0; build the moderation request using the new standalone CSAM types in TypeScript to validate fields at build time.

    Tip: Use classifier signals to prioritize human review rather than as the basis for automated decisions.

    ✓ Verified by official sourceModels: omni-moderation-csam-latestSource: openai-node v7.60.0 sürüm notlarıSource: OpenAI API CSAM Image Classifier rehberi
  • APIOpenAI Agents SDK for Python

    Agents SDK for Python preserves usage from failed Responses calls

    Agents SDK for Python v0.25.0 preserves provider-reported token usage for failed or incomplete Responses calls. This keeps usage and cost observability available for errored or interrupted jobs.

    Usage details

    Who can use it: openai-agents-python v0.25.0 and later; Responses calls.

    How to use it: Upgrade the openai-agents package to v0.25.0 or later and continue using your existing usage and telemetry flow when handling results from failed or incomplete calls.

    Tip: In your error-handling code, write usage information to the same telemetry record as normal successful responses.

  • CapabilityOpenAI Agents SDK for Python demo REPL

    Agents SDK for Python adds per-tool-call approval in the demo REPL

    The v0.25.0 demo REPL supports approving or rejecting each individual tool call. This changes the demo REPL interaction, separately from the managed approval feature previously recorded for the general runtime.

    Usage details

    Who can use it: The demo REPL example in openai-agents-python v0.25.0 and later.

    How to use it: Upgrade the package to v0.25.0, run the demo REPL, and when a tool call is requested, respond to that call with approve or reject.

    Tip: Use the demo REPL approval flow to try out user-approval UX, not as a production authorization policy.

  • SettingOpenAI Agents SDK for Python

    Agents SDK for Python no longer loads startup files in internal sandbox processes

    v0.25.0 prevents workspace startup files from being loaded during internal sandbox processes. This better isolates sandbox commands from unexpected side effects caused by the user’s shell startup settings.

    Usage details

    Who can use it: Sandbox processes in openai-agents-python v0.25.0 and later.

    How to use it: Upgrade the SDK to v0.25.0; explicitly define required environment variables and startup settings in the runtime configuration instead of relying on shell files in the sandbox environment.

    Tip: For reproducible sandbox runs, specify the environment explicitly and avoid tests that depend on the local shell profile.

  • CapabilityOpenAI Agents SDK for Python

    Agents SDK for Python correctly handles explicit interruptions with automatic VAD

    v0.25.0 fixes handling of explicit user interruption requests when automatic VAD is enabled in Realtime streams. This makes manually interrupting a response during a conversation more consistent.

    Usage details

    Who can use it: Realtime streams in openai-agents-python v0.25.0 and later.

    How to use it: Upgrade the SDK to v0.25.0; send an explicit interruption in a Realtime session with automatic VAD enabled and verify that the current response stops.

    Tip: In voice agent tests, test natural VAD interruptions and explicit user-initiated interruptions separately.

  • CapabilityOpenAI Agents SDK for Python

    Agents SDK for Python sessions preserve pending calls and retry history

    v0.25.0 preserves pending asynchronous calls and pending input history between stream retries in checkpoint/session records. This reliability fix is intended to prevent work or the end user’s latest input from being silently lost in resumed sessions.

    Usage details

    Who can use it: openai-agents-python v0.25.0 and later; flows that use session/checkpoint.

    How to use it: Upgrade the SDK to v0.25.0; add regression tests for pending calls and input history in flows resumed with session/checkpoint and using streamed retry.

    Tip: Version checkpoints in production and verify that pending calls are deduplicated and resumed in the correct order after a restart.

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 · 23

  • Plugin / IntegrationGitHub Copilot

    Gemini 4 Neon reaches GA in GitHub Copilot

    GitHub Copilot is making Gemini 4 Neon generally available for agentic coding, multi-step tool use, and knowledge work. The model is rolling out gradually across VS Code, Visual Studio, Copilot CLI, coding agent, github.com, mobile apps, JetBrains, Xcode, and Eclipse.

    Usage details

    Who can use it: Copilot Pro+, Max, Business, and Enterprise; Business/Enterprise administrators can manage access through model policy.

    How to use it: Select Gemini 4 Neon from the Copilot model picker; if you use Business or Enterprise, enable the model under Copilot settings > model policy.

    Tip: For code changes with fewer unwanted edits, first evaluate the model on a narrowly scoped coding-agent task.

    ✓ Verified by official sourceModels: Gemini 4 NeonSource: GitHub Changelog: Gemini 4 Neon in GitHub Copilot
  • ModelGemini 4 Neon

    Gemini 4 Neon reaches GA in the API and Gemini app

    Gemini 4 Neon is now available in the Gemini API, Google AI Studio, Gemini for business, Antigravity, and the Gemini app. The model offers a 1 million-token context window, long decode, and low/medium/high/xhigh thinking levels; API pricing is $3 per 1 million input tokens and $15 per 1 million output tokens.

    Usage details

    Who can use it: Paid tier in the API; Pro and Ultra in the Gemini app; paid Pro/Ultra in Antigravity; global, us, and eu multi-regions in Gemini Enterprise, with admin enablement required in the app.

    How to use it: Select the gemini-4-neon model in the AI Studio model picker or in a Gemini API call; a Gemini Enterprise administrator must enable model availability, and in Antigravity, add it to the authorized model list.

    Tip: Measure the quality-cost tradeoff by choosing high for routine tasks and xhigh for deeper, multi-step coding and analysis work.

  • APIGemini API

    Office files and content slicing come to the Gemini API

    The Gemini API can now directly accept Word, PowerPoint, Excel, and OpenDocument files. In prompts and function response parts, you can select pages, worksheets, cell ranges, time ranges, or content layers so that only the necessary content is processed and billed.

    Usage details

    Who can use it: Gemini API developers; document, video, and audio understanding workflows.

    How to use it: Provide an Office/OpenDocument file in a prompt or function response; use the slicing options in the Document, Video, or Audio understanding guides to narrow the scope by page, tab, cell, time, or layer.

    Tip: For large files, start by slicing only the relevant page or time range to reduce unnecessary tokens and latency.

  • FeatureGemini Enterprise Canvas

    Gemini Enterprise Canvas reaches GA for interactive web app generation

    Gemini Enterprise Canvas can now create single-file, interactive HTML-based web apps, dashboards, and prototypes without writing code. Users can create, preview, edit, export, and share the output.

    Usage details

    Who can use it: Gemini Enterprise; GA. The web app feature must be enabled by an administrator.

    How to use it: Describe an app, dashboard, or prototype in a Gemini Enterprise chat; manage the Canvas output with the Preview, Edit, Export, or Share options.

    Tip: Specify the screen layout, interactions, and export purpose in your first prompt to get a single-file output with fewer revisions.

  • ModelProject Genie

    Genie 4 opens access in the Project Genie experience

    Google has made its Genie 4 world model available to eligible Google AI Ultra subscribers in the new Project Genie experience. Users can modify worlds with real-time prompts, use native audio, run sessions of up to three minutes in Genie 3, and use 360-degree panorama grounding.

    Usage details

    Who can use it: Eligible Google AI Ultra subscribers can try it today; Google AI Pro members can join the waitlist from the Project Genie page.

    How to use it: Go to labs.google/projectgenie; if you’re an Ultra subscriber, start the experience, or if you’re a Pro user, join the waitlist.

    Tip: First compare Genie 3’s longer sessions and panorama grounding, then try real-time audio world editing with Genie 4.

    ✓ Verified by official sourceModels: Genie 4, Genie 3Source: Google DeepMind: Genie 4 kullanıma açıldı
  • AgentStitch Loop

    Stitch Loop opens a waitlist for its always-on product agent

    Stitch Loop was introduced as a long-running product agent designed to connect to a codebase and live metrics, generate recommendations, write code in an isolated sandbox, and open GitHub pull requests for approval. Its safety loop is powered by Gemini 4 Carbon and CodeMender; no PR is merged without user review.

    Usage details

    Who can use it: Global waitlist; priority access for Google AI Ultra users.

    How to use it: Join at stitch.google.com/loop/waitlist; when access arrives, connect GitHub and product metrics, then choose a security, latency, or accessibility loop.

    Tip: Start the first loop with an easy-to-revert latency or accessibility goal, and manually review every PR.

    ✓ Verified by official sourceModels: Gemini 4 CarbonSource: Google Labs: Stitch Loop tanıtımı
  • Plugin / IntegrationDataflow / Data Agent Kit

    Dataflow adds a remote MCP server to Data Agent Kit

    Dataflow’s Public Preview remote MCP server is now integrated with Data Agent Kit. The server exposes regional Dataflow jobs and steps to agents through the get_job, list_jobs, and list_job_steps tools.

    Usage details

    Who can use it: Dataflow and Data Agent Kit users; Public Preview.

    How to use it: Enable the Dataflow remote MCP server and connect it to the agent; call the get_job, list_jobs, and list_job_steps tools in prompts.

    Tip: Keep monitoring prompts narrowly scoped by giving the agent access only to the necessary regions and job IDs.

  • AvailabilityGemini App

    Gemini app comes to low-memory Android Go devices

    The Gemini app has been rearchitected to run on eligible Android Go edition devices. On these devices, users can access Gemini Live, camera-based note/form understanding, and image generation with Nano Banana.

    Usage details

    Who can use it: Global rollout; eligible Android Go edition phones and, soon, devices from partners such as Transsion.

    How to use it: Download the Gemini app from Google Play or press and hold the power button on a supported phone; the device manufacturer may preinstall the app.

    Tip: On a low-memory device, try Gemini Live or short camera queries first; close background apps for longer generations.

  • Plugin / IntegrationGoogle Antigravity

    Robotics plugin for Gemini 4 Hydrogen comes to Antigravity

    The new Robotics plugin in Google Antigravity brings together robot setup, camera discovery and calibration, MuJoCo simulation, perception, and closed-loop control skills in one workspace. Developers can create control code and real-world deployable applications using natural language.

    Usage details

    Who can use it: Antigravity for Google AI Pro or Ultra subscribers; the Robotics plugin is rolling out.

    How to use it: Open the Customizations section in Antigravity, find and install Robotics on the Build With Google shelf, and select Gemini 4 Hydrogen.

    Tip: First verify reach and grasp movements in simulation; then transfer the same control flow to the physical robot.

  • FeatureGoogle AI Studio Build

    Google AI Studio can generate full-stack Android apps

    Google AI Studio Build can now create Android apps from a single prompt, with Kotlin, Cloud Firestore persistent storage, Google Sign-In, and Gemini calls protected by Firebase AI Logic/App Check. Generated apps can be published to a personal Google Play library and installed on a phone.

    Usage details

    Who can use it: Android Build workflow in Google AI Studio; according to the article, the features began rolling out the same day.

    How to use it: Open Google AI Studio Build, select the Android chip at the top, and describe the app; approve backend creation, then publish it to your personal Google Play library from the publish panel.

    Tip: Explicitly specify the data model, sign-in, and Gemini feature in the first prompt; the Build agent will set up the required Firebase components more accurately.

    ✓ Verified by official sourceModels: Gemini 4 Carbon, Gemini 4 HydrogenSource: Google Developers Blog: Gemini ile Google AI Studio’da tam yığın Android uygulamaları
  • SettingModel Armor

    Custom organization policies for Model Armor FloorSetting reach GA

    Organizations can now use custom organization policies to ensure that updated Model Armor FloorSetting resources enforce mandatory filters for responsible AI, prompt injection, and jailbreak detection. The feature is generally available.

    Usage details

    Who can use it: Google Cloud organizations using Model Armor; GA.

    How to use it: Use the custom organization policy workflow in the Model Armor access control documentation to define mandatory filter constraints for FloorSetting, then apply the policy to the project or organization.

    Tip: Lock mandatory filters in a central policy to prevent teams from weakening protections through individual FloorSetting updates.

  • CapabilityTasks in Gemini

    Tasks in Gemini can call US businesses on your behalf

    Gemini can now call multiple US businesses on your behalf about tire availability, service appointments, or price quotes, and return a call summary and full transcript. Calls do not begin until you approve the call plan; Gemini identifies itself as an AI assistant.

    Usage details

    Who can use it: Rolling out in the US; calling is not available in Indiana, Louisiana, Minnesota, Montana, or Nebraska.

    How to use it: Switch the Gemini home screen to Task, enter the businesses to call, a time window, and your questions; approve the plan Gemini presents before calling.

    Tip: Include business hours, preferred dates, and any sensitive information that must not be shared in your call request from the start.

  • Pricing / PlanTasks in Gemini

    Tasks in Gemini is now free and has expanded to six countries

    Running on Gemini 4 Hydrogen, Tasks in Gemini is becoming available for free to hundreds of millions of users in the US, India, Canada, Australia, Japan, and South Korea. The feature handles multistep tasks and recurring routines in the background.

    Usage details

    Who can use it: Free rollout for personal accounts belonging to people aged 18 and over in the six listed countries. Most advanced workflows are starting in the US.

    How to use it: Open the Gemini app, switch the Chat dropdown on the home screen to Task, and describe the task in natural language.

    Tip: For your first try, choose a small routine with an easily verifiable result, such as adding an item to your calendar or creating a morning summary.

  • Plugin / IntegrationTasks in Gemini

    Tasks in Gemini can be managed through Google Messages and iMessage

    Users can now assign tasks to Gemini through Google Messages or iMessage, send a photo to add a date to their calendar, and guide an ongoing task with a reply. Gemini still requires one-tap confirmation for critical actions.

    Usage details

    Who can use it: English; requires a personal account in the U.S. for users over 18, a verified U.S. phone number, and an eligible carrier.

    How to use it: On Android or iOS, go to Gemini > Settings > Channels > Text with Gemini > Set up and verify your phone; alternatively, use gemini.google.com/channels.

    Tip: Instead of giving a long task in one message, narrow down the result step by step with new instructions as it works.

  • Plugin / IntegrationTasks in Gemini

    Approved automatic payments with Google Pay in Tasks in Gemini

    Gemini can now monitor stock and price drops, reorder household products within a budget limit, and prepare payment with Google Pay. A purchase is not completed without a single tap and device approval after you review the itemized summary.

    Usage details

    Who can use it: In English; rolling out on the web, Android, and iOS for personal Google Accounts belonging to people aged 18 and over in the US.

    How to use it: When starting a new Task, specify the product and maximum price; review the cart and summary prepared by Gemini, then provide 1-tap and device approval.

    Tip: State the budget limit clearly for each product, and check the quantity and seller on the summary screen before purchasing.

  • Plugin / IntegrationTasks in Gemini

    Tasks in Gemini can be started by forwarding email

    Users can now start tasks by forwarding emails and PDF attachments from Gmail, Apple Mail, or Outlook to gemini@google.com. Gemini sends the result back in a private 1:1 reply with a review-and-approve link.

    Usage details

    Who can use it: In English; rolling out to personal Google Accounts for people aged 18 and over in the US.

    How to use it: Forward the email thread or PDF attachment to gemini@google.com; open and approve the result using the link in the private reply.

    Tip: When sending multiple school, bill, or renewal emails in one task, clearly state which calendar or reminder actions you want.

  • CapabilityGemini API

    Gemini 3.5 Transcribe supports voice editing commands

    gemini-3.5-transcribe and gemini-3.5-transcribe-live can now edit, rewrite, and format dictated speech or existing text using voice commands. For existing text, source_text can be used in the API.

    Usage details

    Who can use it: Gemini API developers; standard and Live transcription workflows.

    How to use it: Follow the example in the Audio transcription or Live transcription guide; send new dictation or the text to edit using source_text, and include a voice editing instruction.

    Tip: Give editing commands as short, single-purpose phrases, such as ‘add a heading’ or ‘convert to bullets.’

    ✓ Verified by official sourceModels: gemini-3.5-transcribe, gemini-3.5-transcribe-liveSource: Gemini API sürüm notları: Voice editing in Gemini 3.5 TranscribeSource: Gemini API changelog Markdown: Voice editing
  • AvailabilityGemini for Home

    Gemini for Home gets faster and expands to six new countries

    Gemini for Home now processes everyday requests up to 20 percent faster and is expanding to six new countries. Google’s announcement explicitly names India and Brazil among the new countries; the other four are not listed in this source.

    Usage details

    Who can use it: Gemini for Home; six new countries, including India and Brazil. The complete country list was not provided in the source.

    How to use it: In a supported region, make a regular everyday request to the Google Home or Gemini for Home voice assistant; no additional setup steps were specified.

    Tip: To measure the speed difference, try the same short home command several times before and after the update.

  • FeatureGoogle AI Studio Build

    Google AI Studio adds an HD emulator and visual debugging to the Android workflow

    The Android workflow offers a faster 1200p HD emulator view, real-time audio in the browser, one-click screenshot sharing with the agent, custom fonts, and Android CLI-based project/Gradle operations. These tools speed up visual UI fixes and voice-based app testing with Gemini.

    Usage details

    Who can use it: Rolling out to the Google AI Studio Android Build workflow.

    How to use it: Open the Android emulator; enable HD and audio from the toolbar, attach an image to the agent with Take a screenshot, or request a custom font in your prompt.

    Tip: For UI fixes, attach a screenshot and describe a single visual issue; the agent will reduce unnecessary code changes.

    ✓ Verified by official sourceModels: Gemini 4 Carbon, Gemini 4 HydrogenSource: Google Developers Blog: Gemini ile Google AI Studio’da tam yığın Android uygulamaları
  • AvailabilityGoogle Vids

    Google Vids makes personal avatars free for everyone

    Anyone with a personal Google account can now create a personal avatar with their own appearance and voice in Google Vids. The avatar is created using a QR code, saved to a Vids project, and can be reused in scenes.

    Usage details

    Who can use it: Anyone with a personal Google account; free and available on desktop.

    How to use it: On desktop, open vids.new, start the personal avatar workflow, scan the QR code to create your avatar, and use it in a video prompt.

    Tip: Before saving the avatar, test different scenes and speaking styles to create a neutral baseline version for reuse.

  • AvailabilityGoogle Wallet

    Google Wallet launches on iOS in the U.S.

    Google Wallet is now available to iOS users in the U.S.; it can manage cards, loyalty cards, tickets, receipts, and records of purchases facilitated through Gemini/Google in one place. Tap-to-pay is not yet supported on iOS.

    Usage details

    Who can use it: iOS in the U.S.; Google Wallet app. Tap-to-pay is not supported on iOS.

    How to use it: Install Google Wallet from the U.S. App Store, sign in with your Google account, and add your cards or passes.

    Tip: Keep notifications and Gmail sync on to track orders and delivery records created with Gemini Tasks.

  • SettingAI Mode in Search

    Temporary threads arrive in Search AI Mode

    AI Mode now has a temporary threads feature; these sessions aren’t saved to history, personalized using past preferences, or used to influence future recommendations. The conversation is removed when the session ends.

    Usage details

    Who can use it: Global rollout in desktop and mobile browsers, in all languages supported by AI Mode.

    How to use it: In AI Mode, click the temporary threads icon in the upper-right corner and enter your query in a new temporary session.

    Tip: Use it for gift, health, or sensitive research queries when you don’t want the temporary session to affect personalization.

  • Plugin / IntegrationTasks in Gemini

    Tasks in Gemini can use portal data with Google Password Manager

    Google’s November 17 announcement shows that, with the user’s permission, Gemini can access service portals using information in Google Password Manager and analyze bills and similar documents. This brings a new credentials and web portal integration to Tasks workflows.

    Usage details

    Who can use it: The advanced Tasks example in the announcement is part of a rollout starting in the U.S.; a broader list of countries was not specified.

    How to use it: Ask Gemini to analyze portal data within a Task; when Gemini requests permission, verify access to Google Password Manager and approve critical steps.

    Tip: Enable portal access only for tasks that need it, and review connected app permissions after the task is complete.

    ✓ Verified by official sourceModels: Gemini 4 HydrogenSource: Google: Gemini’yi sizin için çalıştırın

Meta · Meta AI

Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 2

  • Plugin / IntegrationMuse

    Loblaw becomes Muse’s first Canadian retail connector

    Loblaw’s product catalog has been added to Muse as a partner connector. Users in Canada can plan meals, build a cart and complete checkout within Muse at participating Loblaw brands.

    Usage details

    Who can use it: Muse users in Canada and participating Loblaw brands.

    How to use it: Connect the Loblaw partner connector to your account in Muse; enter your meal or product request, let Muse build your cart and approve the purchase before checkout.

    Tip: Clearly specify allergies and brand preferences; verify the cart, quantities and delivery details yourself before purchasing.

  • AvailabilityMuse

    Muse Voice expands to mobile, web and Mac in Canada

    Following its initial rollout in the US, Muse Voice is beginning to launch in Canada on mobile, web and Mac. In real-time conversations, users can choose audio-only, an avatar or full-screen view.

    Usage details

    Who can use it: Muse users in Canada; mobile, web and Mac platforms; AI glasses are not included in this announcement.

    How to use it: Open Muse and tap the phone icon in the top right; start a conversation, then choose audio-only, avatar or full-screen view.

    Tip: Adjust the level of visual engagement to suit your needs: use audio-only while walking or when your hands are full, and an avatar or full-screen view while developing ideas.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 10

  • SettingKimi Work

    Ask user notifications are not sent while Kimi Work is in focus

    Ask user notifications are now sent only when the app is not in focus. This reduces unnecessary foreground interruptions while actively using the app.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan and operating system coverage were not specified in the official notes.

    How to use it: Update Kimi Work to 3.2.28; follow Ask user prompts in the app while it is in focus, and receive notifications when the app is in the background.

    Tip: Keep the operating system’s notification permission for Kimi Work enabled so you don’t miss notifications for background tasks.

  • CapabilityKimi Work

    Kimi Work shows sources, times, and summaries in global search

    Global search results now include the source, time, and a short summary. Hovering over a result also displays a preview of the surrounding context.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan coverage was not specified in the official notes.

    How to use it: Open global search from the sidebar, run a query, and review the source, time, summary, and hover preview for the results.

    Tip: To distinguish similar results, check the source and time first, then review the context in the hover preview.

  • FeatureKimi Work

    Kimi Work summary panel brings together progress, files, outputs, and scheduled tasks

    The summary panel now shows conversation progress, uploaded files, outputs, and scheduled tasks in one place; you can open files and their containing folders from here.

    Usage details

    Who can use it: Users of Kimi Work 3.2.28; plan and operating system coverage were not specified in the official note.

    How to use it: Open the summary panel in a conversation; open the relevant file or folder from the progress, uploaded files, outputs, or scheduled tasks section.

    Tip: For long-running tasks, check this panel first to find the latest progress and generated files in one view.

  • FeatureKimi Work

    Project-level AGENTS.md editing added through the Kimi Work summary panel

    You can add and edit a project-level AGENTS.md file through the summary panel. These instructions apply to all conversations in the relevant project.

    Usage details

    Who can use it: Project users on Kimi Work 3.2.28; plan coverage was not specified in the official notes.

    How to use it: Open the project’s summary panel, add an AGENTS.md file in the Project instructions section or edit the existing file, then save it.

    Tip: Write recurring project rules and security boundaries in AGENTS.md as short, clear bullet points.

  • FeatureKimi Work

    Kimi Work delivers new presentations as PPTX by default

    The delivery format for presentations newly created with Kimi Work is now PPTX by default.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan and platform distinctions were not specified in the official notes.

    How to use it: Create a new presentation in Kimi Work; if you don’t select another format in the delivery card, the result will be provided as a PPTX.

    Tip: Keep the default PPTX delivery if you plan to edit the presentation in PowerPoint; convert it afterward only if you need a different output format.

  • CapabilityKimi Work

    Kimi Work detects running tasks and warns before an update

    Running tasks are detected before an app update is installed, and users are notified that the update could interrupt them.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan and operating system coverage for the update flow was not specified.

    How to use it: Start a Kimi Work update; if there are running tasks, read the warning and continue with installation after stopping the tasks or accepting the risk.

    Tip: Let long-running background tasks finish or save their outputs before updating.

  • AvailabilityKimi Work

    Enterprise SSO sign-in reported for Kimi Work

    Two secondary reports say that with Kimi Work 3.2.28, enterprise users can sign in with Enterprise SSO using their work email. This is not considered confirmed by Moonshot because the official Kimi 3.2.28 notes do not include this item.

    Usage details

    Who can use it: Enterprise Kimi Work users; Enterprise SSO using work email. Official plan, country, and identity provider coverage were not specified.

    How to use it: On the Kimi Work sign-in screen, select the Enterprise SSO flow using your work email and sign in with your organization’s identity provider; news sources do not provide details on the menu name.

    Tip: For enterprise deployment, first confirm your organization’s SSO domain and the identity provider configuration required for Kimi Work with your administrator.

  • CapabilityKimi Work

    Kimi Work recognizes kimi-design delivery cards and improves previews

    Kimi Work now recognizes kimi-design delivery cards and improves previews of design assets and presentations.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; coverage for use with kimi-design outputs was not detailed in the official notes.

    How to use it: Create a design or presentation output with kimi-design in Kimi Work; open the delivery card in the conversation to use the new preview.

    Tip: Review design deliverables through the card and check the presentation layout in the preview before downloading the file.

  • CapabilityKimi Work

    Kimi Work offers a retry option for background tasks that go unanswered for a long time

    When a background task goes unanswered for a long time, an explicit status message is now displayed along with a retry option.

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan and platform coverage were not specified in the official notes.

    How to use it: Start a background task; if there is no response for a long time, use the Retry option in the explicit status message that appears.

    Tip: Before using Retry repeatedly, check whether the task is waiting on an external system or network.

  • FeatureKimi Work

    Kimi Work combined the web annotation entry under “Annotate/Edit”

    The entry used for web annotation is now combined under the name “Annotate/Edit.”

    Usage details

    Who can use it: Kimi Work 3.2.28 users; plan and platform coverage were not specified in the official notes.

    How to use it: Open a web page in Kimi Work’s built-in browser and select the Annotate/Edit entry to start an annotation or editing flow.

    Tip: Use the same Annotate/Edit entry to choose the right flow when switching between adding annotations and changing content.

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 23

  • Pricing / PlanQoder

    Qoder lowered the Auto, Fast, and Efficient Credit multipliers

    At 10:00 a.m. Singapore time on November 18, 2026, Qoder lowered the Auto multiplier from 0.5x to 0.3x, Fast from 0.8x to 0.6x, and Efficient from 0.3x to 0.1x. Ultimate remained at 2.0x; existing Credit balances were unchanged.

    Usage details

    Who can use it: Applies to individual, Team, Enterprise, and Service Account users across Qoder, Qoder IDE, the JetBrains plugin, Qoder CLI, QoderWake, Cloud Agents, Mobile, and other Qoder products.

    How to use it: Open the model selector in any Qoder product; the current multipliers are applied automatically and displayed in the selector.

    Tip: Allocate the same Credit balance according to the task by choosing Efficient or Auto for routine work and Ultimate for difficult tasks where errors are costly.

  • AvailabilityQoder

    Qoder launched day-one support for Qwen4-Max

    Qoder added Qwen4-Max, co-developed with Alibaba Qwen, to its product with Day-0 support on November 18. The announcement says Qoder Harness turns model context into actions in PowerPoint, Excel, and Word tasks.

    Usage details

    Who can use it: Qoder; the announcement specifies PowerPoint, Excel, and Word workflows with Qoder Harness. No plan or regional restrictions were disclosed.

    How to use it: Upgrade Qoder to the latest version, select Qwen4-Max in the model selector, and start the task.

    Tip: For long documents and office files, clearly specify the expected output format and evaluation criteria before prompting the model.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Qwen4-Max: Day-0 support in Qoder
  • AgentQoderWake

    QoderWake enables proactive group chat replies without @

    QoderWake Digital Workers can now evaluate relevant group chat requests without an explicit @Waker. The new workflow uses message collection, rapid prescreening with JEV, responsibility and ownership decisions by Router, and execution in a separate Work context; results are proactively sent to the relevant group.

    Usage details

    Who can use it: Available for QoderWake Digital Workers in group chats with configured responsibilities. Execution is subject to existing queue and concurrency limits; this does not mean unlimited parallelism.

    How to use it: In the QoderWake console, add, edit, or delete Waker's proactive response policy in natural language; you can also perform the same actions by explicitly asking Waker in a group chat.

    Tip: Include both the requests the policy should accept and cases that look similar but should be ignored; this helps reduce incorrect ownership claims.

  • SettingQoder

    Qoder changed the default thinking effort for some models from Maximum to High

    Qoder changed the default thinking effort from Maximum to High for GLM-5.3, Kimi-K3.1-Pro, Kimi-K2.8-Preview, and DeepSeek-Flash. GLM-5.3-Flash and DeepSeek-V4-Pro remained at Maximum; Low, High, or Maximum can be selected manually for all these models.

    Usage details

    Who can use it: The relevant models in Qoder's model selector; plan or regional availability was not specified.

    How to use it: Open the model selector below the input box, go to Manage models, and set the relevant model's thinking effort to Low, High, or Maximum.

    Tip: Use High for everyday development, Low for small, well-defined edits, and Maximum for difficult debugging and thorough verification.

    ✓ Verified by official sourceModels: GLM-5.3, Kimi-K3.1-Pro, Kimi-K2.8-Preview, DeepSeek-Flash, GLM-5.3-Flash, DeepSeek-V4-ProSource: What Is Model Thinking Effort? Should You Choose Low, High, or Maximum?
  • CapabilityQoder CLI

    Qoder CLI reduced input token usage for repeated calls to the same Skill

    Qoder CLI 1.1.92 reduced input token usage when the same Skill is called repeatedly within a session.

    Usage details

    Who can use it: Qoder CLI 1.1.92 and later.

    How to use it: Upgrade Qoder CLI to version 1.1.92 or later; call the Skill again as usual within the same session.

    Tip: Keeping repeated operations in a session under the same Skill makes better use of this optimization.

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

    Qoder CLI distinguishes heredoc content from dangerous command warnings

    Qoder CLI 1.1.92 fixed an issue where heredoc content containing text such as rm and backticks was incorrectly flagged as a dangerous command.

    Usage details

    Who can use it: Qoder CLI 1.1.92 and later; particularly terminal workflows that use heredocs.

    How to use it: Update the CLI to version 1.1.92 or later and run heredoc commands as usual; their content should no longer be flagged due to a false match.

    Tip: Still review commands that will actually execute inside a heredoc; the fix targets only false-positive classifications.

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

    Qoder CLI improved compaction retry reliability

    Qoder CLI 1.1.92 fixed errors in context compaction retries and improved compaction stability in long sessions.

    Usage details

    Who can use it: Qoder CLI 1.1.92 and later; no plan or operating system restrictions specified.

    How to use it: Update Qoder CLI to version 1.1.92 or later; context compaction will automatically use the new stability flow in long sessions.

    Tip: For long tasks, check whether the last user instruction and critical file changes are preserved after compaction.

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

    Qoder CLI now matches the model in transcript records to the model actually used

    Qoder CLI 1.1.92 fixed an issue where the model recorded in the transcript could differ from the model actually used.

    Usage details

    Who can use it: Qoder CLI 1.1.92 and later.

    How to use it: Update the CLI; in new sessions, open the transcript record to check the selected and used model information.

    Tip: When comparing models, verify the CLI version before relying on the model name in the transcript.

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

    Qoder CLI fixed missing Assistant Credit information in transcript records

    Qoder CLI 1.1.92 fixed an issue where Credit information for assistant messages was missing from transcript records.

    Usage details

    Who can use it: Qoder CLI 1.1.92 and later.

    How to use it: Update the CLI and review new transcript records; the Credit field for assistant messages will be written again in supported flows.

    Tip: When tracking costs, compare old and new transcript records with the version difference in mind.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • Plugin / IntegrationQoder CLI

    Enterprise administrators can distribute Qoder CLI Skills and MCPs to members

    Qoder CLI 1.1.91 added support for enterprise administrators to distribute standalone Skill and MCP configurations to members.

    Usage details

    Who can use it: Enterprise Qoder CLI deployments; administrators and organization members. Detailed licensing limits were not specified.

    How to use it: As an administrator, assign standalone Skill and MCP packages to members through the organization's Qoder CLI administration deployment; members can open the CLI with the latest version and use the resources.

    Tip: Roll out distribution in stages, and apply least-privilege access controls to MCP endpoints shared within the organization.

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

    Qoder CLI no longer unnecessarily restores old context after /rewind

    Qoder CLI 1.1.91 fixed an issue where using `/rewind` after auto-compaction restored all pre-compaction history and started another unnecessary compaction.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; long sessions using auto-compaction and `/rewind`.

    How to use it: Use `/rewind` as usual in a long session; the CLI checks context usage to prevent unnecessary restoration of the full history.

    Tip: Check the context indicator after rewinding; restoring only the necessary checkpoint uses less context.

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

    Qoder CLI Auto permission mode requests approval only for high-risk commands

    Qoder CLI 1.1.91 changed the policy so that, in Auto permission mode, model approval is triggered only when a high-risk command is matched.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; sessions using Auto permission mode.

    How to use it: Select Auto permission mode in the CLI; approval prompts appear for high-risk commands, while unnecessary model approval is reduced for other matches.

    Tip: Even when using Auto mode for commands with significant operating system or repository impact, review approval prompts carefully.

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

    Qoder CLI no longer completely disables @ completion when it cannot build a file index

    Qoder CLI 1.1.91 fixed an issue that caused `@` completion to fail completely when a project directory could not build a file index.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; project directories where a file index cannot be built.

    How to use it: Update the CLI and try completing a file or folder with `@` in the prompt box; the prompt flow should no longer crash if there is an indexing issue.

    Tip: If indexing issues persist in large or access-restricted repositories, use explicit file paths instead.

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

    Qoder CLI WebFetch no longer asks again after selecting Always allow

    Qoder CLI 1.1.91 fixed an issue where the same permission prompt appeared again after selecting Always allow for WebFetch.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; sessions using WebFetch.

    How to use it: Select Always allow in the WebFetch permission prompt; the updated CLI should not ask for permission again for the same scope.

    Tip: Select Always allow only for WebFetch scopes you trust, and review permissions regularly.

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

    Qoder CLI cleans up processes from canceled Git probes on Windows

    Qoder CLI 1.1.91 fixed an issue where Git processes remained running when a session's Git context probe timed out or was canceled on Windows.

    Usage details

    Who can use it: Qoder CLI sessions using Git context on Windows; CLI 1.1.91 and later.

    How to use it: Update the CLI and start a session in a Git repository; processes are cleaned up automatically after cancellation or a timeout.

    Tip: Monitor CPU usage during long-running repository probes; if processes remain from an older version, fully close and reopen the CLI.

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

    Qoder CLI preserves BYOK thinking and context options

    Qoder CLI 1.1.91 fixed an issue where thinking effort, disable thinking, and context window options were lost for BYOK models.

    Usage details

    Who can use it: Qoder CLI users with BYOK models; CLI 1.1.91 and later.

    How to use it: Select and save the thinking effort, disable thinking, or context window value in BYOK model settings; the same settings will be preserved for subsequent use.

    Tip: Assume settings may be model-specific when switching models, and check the values again before the first request in a new session.

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

    Qoder CLI trims history and retries compression on input-too-long errors

    Qoder CLI 1.1.91 trims the oldest history and retries context compression if the model service returns input-too-long during a long conversation.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; long conversations and sessions approaching the context limit.

    How to use it: Update the CLI and continue the long session; history trimming and compression run automatically when input-too-long occurs.

    Tip: In very long sessions, restating critical requirements in interim summaries helps reduce context loss after automatic trimming.

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

    Qoder CLI can complete Windows installation in restricted PowerShell

    Qoder CLI 1.1.91 fixed a package extraction issue in PowerShell Constrained Language Mode environments such as those using AppLocker or WDAC. When needed, installation falls back to the built-in Windows tar.exe.

    Usage details

    Who can use it: Machines using PowerShell Constrained Language Mode on Windows, especially corporate environments with AppLocker or WDAC enforced.

    How to use it: Run the official Windows installer for Qoder CLI 1.1.91 or later; the tar.exe fallback activates automatically in compatible environments.

    Tip: For enterprise deployment, first verify the installation on a test machine with Constrained Language Mode and policy restrictions.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes
  • Plugin / IntegrationQoder CLI

    Qoder CLI keeps MCP tools available when a list refresh fails

    Qoder CLI 1.1.91 fixed an issue that made tools unavailable when refreshing the MCP tool list failed.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; sessions using MCP connections.

    How to use it: Update the CLI, configure your MCP connection, and refresh the MCP tool list during a session; a refresh failure should not permanently hide the tools.

    Tip: If the issue persists, check the connection status and tool list again in the `/mcp` view.

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

    Qoder CLI fixes duplicate session content in /export output

    Qoder CLI 1.1.91 fixed an issue that caused session content to be exported twice when `/export` was used in some terminals.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; especially some terminal environments.

    How to use it: Use the `/export` command in a session; the updated CLI should not produce duplicate content.

    Tip: Before archiving, quickly check that headings and messages appear only once in the exported file.

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

    Qoder CLI displays the input cursor in terminals with limited color support

    Qoder CLI 1.1.91 fixed an issue where the cursor was not visible in the main input box in terminals with limited color support, such as TERM=xterm under GNU screen.

    Usage details

    Who can use it: Terminals with limited color support; especially GNU screen and environments similar to TERM=xterm.

    How to use it: Upgrade the CLI to version 1.1.91 or later and run it in the current terminal session; no additional configuration is required.

    Tip: If the cursor is not visible, first check the `TERM` value and terminal multiplexer configuration.

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

    Qoder CLI clears leftover colors when switching themes

    Qoder CLI 1.1.91 clears leftover colors after switching between light and dark terminal themes.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; light/dark terminal themes.

    How to use it: Update the CLI and switch the terminal theme; no additional configuration is needed.

    Tip: If the colors still look wrong, resizing the terminal or starting a new session is a quick workaround.

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

    Qoder CLI can open the full URL when it wraps across lines

    Qoder CLI 1.1.91 fixed an issue where clicking a long URL that wrapped across lines in the session stream did not open the full address.

    Usage details

    Who can use it: Qoder CLI 1.1.91 and later; long URLs in the session stream.

    How to use it: Click a URL split across lines in the session stream; the CLI will now try to open the full address.

    Tip: For untrusted links, check the domain and full path before opening.

    ✓ Verified by official sourceSource: Qoder CLI Release Notes

Watchlist

Muse Wonder Full

50
Meta · Meta AI

Tied with Gemini 4 Argon at 50 points, but dropped out of the TOP 10 because it ranks 11th in the Artificial Analysis table. Meta’s strongest Muse model; no public API. Muse gained its first retail connector in Canada (Loblaw) and Muse Voice.

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. Kimi Work 3.2.28 added project-level AGENTS.md editing, a unified summary panel, and default PPTX delivery. Qoder lowered Kimi-K3.1-Pro’s default thinking effort from Maximum to High.

Muse Wonder

44
Meta · Meta AI

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

Qwen4 Max

43
Alibaba · Qwen

Released on November 16, it is 14th with 43 points. `qwen4-max` in Model Studio (1M context, Singapore pricing $1,50 / $4,50), QwenCloud Token Plan, QwenWork, QwenStudio, and OpenRouter (`qwen/qwen4-max`). Qoder added day-one support for Qwen4-Max on November 18. Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

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

Muse Ember

36
Meta · Meta AI

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

Sources