YZ·RADARyz.mikail.net
Daily edition · 10 November 2026

Claude Fable 5.2 tops the ranking with 65 points; Meta’s $0,10 Muse Ember

After five days of inactivity, the TOP 10 changed. Claude Fable 5.2, released by Anthropic on November 9, went straight to the top of the Artificial Analysis Intelligence Index v5.0.1 with 65 points, beating Gemini 4 Carbon (60) by 5 points. The model’s xhigh and high settings also score 62 points, putting it ahead of the list without choosing its most expensive setting. With this debut, eight models—from Gemini 4 Carbon to Muse Wonder Full—each dropped one place, with no change in their scores. Claude Fable 5.1 was marked “deprecated” on Artificial Analysis and removed from the current list (its score fell to 49), dropping out of the TOP 10; Kimi K3.1 Pro remains in 10th place with 46 points. The latest accessible snapshot of the LMArena text ranking is still dated October 27, and Fable 5.2 is not yet listed there. Meta’s new small model, Muse Ember, entered the watchlist with 36 points; DeepSeek V4.1 Flash 1017 (33) remained outside the list.

Here are the details for those who want to use Fable 5.2: 1 million tokens of context, 128 thousand tokens of output, and continuous adaptive thinking; pricing is $8 per 1M tokens for input and $40 for output (cheaper than Fable 5.1’s $10 / $50). The model is available under the `claude-fable-5-2` ID on the Claude API, Claude Platform, Amazon Bedrock, Google Cloud Agent Platform (GA), and Microsoft Foundry. Its sibling model, Claude Mythos 5.2, shares the same base model but is available only to organizations accepted into Anthropic’s cybersecurity and life sciences validation programs. In the API, `speed: "fast"` increases output speed by up to 2.5x ($16 / $80). There are two breaking changes to note when migrating: `computer_toolset_20260801` is required for computer use (the old `computer_20251124` returns a 400 error outside Bedrock), and prefix checks are now required for all accounts in applications that resend thinking blocks from previous turns. Both models are considered Covered Models: data is retained for at least 30 days and automatic safety reviews apply; zero data retention is available only with Anthropic’s approval. Claude Code 2.1.321 made Fable the default model, and 2.1.322 added fast mode support on Vertex (global, US, EU). From November 9–16, Pro and Team standard seats can use up to 50% of their weekly Fable limits at no extra cost and receive one limit reset; Max, Ultra, and premium seats get between one and three resets.

Meta released its smallest and cheapest model, Muse Ember, on November 9: $0,10 / $0,50, 1M context, up to 512 thousand output tokens, and text, image, audio, and video input. It scores 36 on Artificial Analysis, matching Claude Haiku 5.5; Meta says it does so at about one-third the task cost (around $0,14 per task). The model is available on Meta Model API as `muse-ember-1.0` (with an OpenAI-compatible endpoint and Responses API support) and on all Muse Code plans; Muse Wonder, Ember, and Spark are in preview on Google Agent Platform. Meta also expanded AI age estimation on Instagram worldwide. On the OpenAI front, ChatGPT Enterprise became available for purchase through AWS Marketplace. Agents SDK prereleases (Python v0.24.0-alpha.7, JavaScript v0.22.0-alpha.2) added FileSearch, CodeInterpreter, and ImageGeneration tools for native agents, advanced web search, A2A, RunBudget, and application-owned session storage. There are also breaking changes: the default reasoning effort for gpt-6-astra and gpt-6.1-sol increased from `low` to `medium`, MCP approval names changed, and `SDKRunner` was removed in JavaScript. Go SDK v3.98.0 added asynchronous function tools and a prompt-cache TTL setting. In Google AI Studio, Maps Platform can now be connected to the Build screen with one click; the trial environment includes 100 free calls per API per day.

Other companies mostly updated agent and developer tools. Qwen Code 0.27.1 added optional Mem0 personal memory for Managed Agent sessions, model and context statistics, channel delivery for scheduled tasks, and self-installation for Skills/MCP; Alibaba Cloud DevOps introduced Review Memory, which remembers team rules during code reviews. Kimi Code now lets users save permission rules persistently at the user and project levels, and select between same-named Skill variants in multi-directory projects. Tencent released its 323-million-parameter 2D PDE model HyPDE under the Apache-2.0 license; six new agent CLIs, including Kimi and DeepSeek Harness, were added to LoopForge, while `score --run --queue` was removed from teamai-cli. xAI reset Grok Bot’s weekly usage limits for everyone on November 9. No new changes have appeared in official sources for DeepSeek, Z.ai, or Xiaomi since November 8. A total of 158 new developments were added to the archive today. To-dos: if you use the Claude API, test your computer use and thinking flows before switching to Fable 5.2; if you’re a Pro/Team subscriber, use your Fable bonus by November 16; if you run many small agent tasks, compare costs with Muse Ember.

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 · ClaudeNEW
    65pts

    Released on November 9 and went straight to the top; the xhigh and high settings also score 62 points. 1M context, 128K output, $8 / $40; available in Claude apps, the API, Bedrock, Google Cloud, and Microsoft Foundry (`claude-fable-5-2`).Released: 9 November 2026

  2. 02

    Gemini 4 Carbon

    Google · Gemini▼ 1
    60pts

    Its score is unchanged, but it fell to second place with the arrival of Fable 5.2. Access remains limited to selected cyber defenders and Fairwind partners; general API and app access is not yet available.Released: 28 October 2026

  3. 03

    Claude Opus 5.5

    Anthropic · Claude▼ 1
    59pts

    Anthropic’s flagship Opus model; available on Claude Pro/Max/Team/Enterprise and the API. Claude Code 2.1.322 added fast mode support on Vertex (global, US, EU).Released: 22 September 2026

  4. 04

    GPT-6 Astra

    OpenAI · ChatGPT▼ 1
    57pts

    OpenAI’s most powerful GPT-6 model. In Agents SDK Python v0.24.0-alpha.7, the default reasoning effort for this model increased from `low` to `medium`.Released: 3 September 2026

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT▼ 1
    55pts

    The default model on ChatGPT paid plans; API pricing is $2 / $10. The default reasoning effort became `medium` in the Agents SDK alpha release.Released: 29 September 2026

  6. 06

    Claude Sonnet 5.5

    Anthropic · Claude▼ 1
    55pts

    The most powerful model available on the free claude.ai plan; tied with GPT-6.1 Sol at 55 points.Released: 28 September 2026

  7. 07

    GPT-6.1 Terra

    OpenAI · ChatGPT▼ 1
    51pts

    Released on October 26 in ChatGPT Work, Codex, and the API; focused on image and computer use. The admin preview for Dots agents runs through November 18.Released: 26 October 2026

  8. 08

    Gemini 4 Argon

    Google · Gemini▼ 1
    50pts

    Available with Google AI Pro/Ultra and the paid API; the $2 / $10 introductory price runs through November 30. Long output (long decode) of up to 1 million tokens in a single request became GA in the Interactions API.Released: 30 September 2026

  9. 09

    Muse Wonder Full

    Meta · Meta AI▼ 1
    50pts

    Meta’s most powerful Muse model; no public API. The first small model trained with it, Muse Ember, launched on November 9.Released: 12 October 2026

  10. 10

    Kimi K3.1 Pro

    Moonshot AI · Kimi=
    46pts

    API $2 / $10; open weights on Hugging Face (2,8T total / 104B active, 1M context). Kimi Code now lets users save permission rules persistently at the user and project levels.Released: 23 October 2026

Key signals

  • Claude Fable 5.2 entered the TOP 10 directly at No. 1 with 65 points; eight models, including Gemini 4 Carbon (60), each dropped one place, with scores unchanged.
  • Claude Fable 5.1 was deprecated on Artificial Analysis (49 points) and dropped out of the TOP 10; Kimi K3.1 Pro remained in 10th place with 46 points.
  • Anthropic: Fable 5.2 costs $8 / $40, less than Fable 5.1; it is available on the API, Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.2 is available only to verified organizations.
  • Meta: Muse Ember launched. $0,10 / $0,50, 1M context, 36 points on Artificial Analysis; available on Meta Model API as `muse-ember-1.0` and on all Muse Code plans.
  • OpenAI: ChatGPT Enterprise is on AWS Marketplace. Agents SDK prereleases added hosted tools, A2A, and RunBudget; default reasoning effort is now `medium`.
  • 158 new developments: Anthropic 54, OpenAI 32, Alibaba 26, Tencent 19, Moonshot 15, Google 6, Meta 4, xAI 2. No new changes at DeepSeek, Z.ai, or Xiaomi.

Worth trying today

01

Check the two breaking changes before switching to Fable 5.2

Set the model to `claude-fable-5-2` in your API request. If you use computer use, replace the old `computer_20251124` definition with `computer_toolset_20260801`; otherwise, you’ll get a 400 error on every surface except Bedrock. If you resend thinking blocks from previous turns, don’t change the system prompt, tools, or message prefix; if you need to, try the beta `thinking.block_binding.prefix_mismatch_behavior: "drop_block"` option. First benchmark with the default `high` effort, and move to `max` only if you see a quality difference in evals. Add `speed: "fast"` for latency-sensitive flows; it costs $16 / $80 and does not work with the Batch API.

Anthropic · Claude
02

Use your Fable bonus before November 16 if you’re a Pro or Team subscriber

From November 9–16, Pro and Team standard seats can use Fable models up to 50% of their weekly limits at no extra cost and receive one Fable limit reset. In Claude, go to Settings > Usage > Limit resets and select “Reset for free.” The reset doesn’t increase your limit; it only restores usage spent on Fable, so save it for a genuinely intensive task. If you update Claude Code to 2.1.321 or later, Fable 5.2 becomes the default; to switch back to the older model, enter `/model claude-fable-5-1`.

Anthropic · Claude
03

Move recurring sub-agent tasks to Muse Ember and measure the cost

Create an API key through dev.meta.ai. In your existing OpenAI-compatible client, set `base_url` to `https://api.meta.ai/v1` and the model to `muse-ember-1.0`. Keep planning and supervision on a larger model (Muse Wonder, Claude, or GPT), and assign large numbers of well-defined tasks (file conversion, document summarization, batch image labeling) to parallel Ember agents. In Meta’s example, a job involving 10,000 photos cost about $1,33 instead of about $20. Note that the price triples ($0,30 / $1,50) for prompts exceeding 200 thousand tokens.

Meta · Meta AI
04

Prototype your map-based app idea in minutes with AI Studio

Go to ai.studio/build and describe your app idea to Gemini (for example, “a page that shows training participants the nearest café and public transit route”). Enable Google Maps Platform with one click or connect your existing Google Cloud project; Maps, Places, Routes, and Environment tools are available. The trial environment doesn’t require a credit card and includes 100 calls per API per day. When moving to production, enable only the Maps APIs you use and billing.

Google · Gemini

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#3 Claude Opus 5.5#6 Claude Sonnet 5.5Claude Haiku 5.5
Full history (836) →

New in this edition · 54

  • AvailabilityClaude Code

    Claude Code fast mode expands to Vertex regions

    Claude Code 2.1.322 supports fast mode on Vertex in the global, US, and EU regions. If Vertex rejects a fast request, the client falls back to standard speed.

    Usage details

    Who can use it: Claude Code 2.1.322 and Vertex models that support fast mode; global, US, and EU regions.

    How to use it: Update Claude Code to version 2.1.322 or later and use the `/fast` command or the fast mode setting with a supported model.

    Tip: Because quotas and model support on Vertex can vary, monitor the latency and cost of the standard-speed fallback in logs.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2, Claude Opus 5.5Source: Claude Code changelogSource: Claude Code release v2.1.322
  • APIClaude Code

    MCP hook responses over 16 MiB now cause calls to be rejected

    In stdio and HTTP MCP servers, PreToolUse `mcp_tool` hook responses larger than 16 MiB are no longer skipped; the tool call is explicitly rejected. This prevents large hook output from entering the context unchecked.

    Usage details

    Who can use it: Claude Code 2.1.321; stdio and HTTP servers using PreToolUse MCP hooks.

    How to use it: Keep hook responses under 16 MiB; write large data to a file and return the file path to Claude.

    Tip: Check the response size limit in the MCP server in advance and report rejections as meaningful errors.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • CapabilityClaude Code

    NotebookEdit and Edit reliability fixed for network paths

    Fixes address NotebookEdit deleting/overwriting the first cell and Edit changing the wrong match, replacing the entire file with an empty `old_string`, or editing a notebook as if it were plain text on UNC, WSL, and `/net` paths.

    Usage details

    Who can use it: Claude Code 2.1.321; Windows network paths, WSL, and other network filesystems.

    How to use it: Update Claude Code and rerun notebook/file editing workflows on network paths.

    Tip: After writing to a network filesystem, automatically check the diff and the contents of the first and last cells.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • AgentClaude Code

    Background subagent reports are preserved after compaction and restart

    Auto mode no longer loses a subagent report that started before compaction; after a restart, the subagent's message to the main conversation is no longer left unanswered; and a background agent mistakenly thought to have stopped can be restarted. This improves the reliability of long-running, interrupted agent workflows.

    Usage details

    Who can use it: Claude Code 2.1.321; background agent, compaction, and cloud restart scenarios.

    How to use it: In the updated version, start the subagent as usual; if no report arrives, check its status and the option to reopen it in the Agents/Tasks view.

    Tip: For long-running tasks, write subagent reports to the Tasks history as well as relying on the latest notification.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • Plugin / IntegrationClaude in Chrome

    Chrome tools in Claude Code comply with organizational policy

    Claude in Chrome tools provided by Claude Code follow the organization’s Claude in Chrome policy in embedded applications such as the Claude Desktop Code tab. Enterprise tool restrictions now also apply to this embedded use.

    Usage details

    Who can use it: Claude Code 2.1.321, Claude Desktop Code tab, and customers who manage organizational policies.

    How to use it: Have the organization administrator manage the Claude in Chrome policy; after updating Claude Code, check the permissions for embedded tools.

    Tip: Manage permissions for the Chrome extension and Desktop Code tab under the same organizational policy, rather than separately.

    ✓ Verified by official sourceSource: Claude Code changelog
  • ModelClaude Code

    Fable 5.2 added to Claude Code, and Fable is now the default

    Claude Code 2.1.321 added the `claude-fable-5-2` model and made the Fable model the default. Saved Fable 5.1 selections move to the `fable` selection immediately on Pro and Max, and on Team and Enterprise when the organization allows it.

    Usage details

    Who can use it: Claude Code 2.1.321; subject to plan and organizational model permissions.

    How to use it: Update the version; use `/model claude-fable-5-1` to force the old model or `/model claude-fable-5-2` to select the new ID.

    Tip: Before switching models, compare token costs, thinking behavior, and tool compatibility with a short eval on critical projects.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Fable 5.1Source: Claude Code changelogSource: Claude Code release v2.1.321
  • ModelClaude API

    Claude Fable 5.2 and Claude Mythos 5.2 are now available

    Anthropic has made Claude Fable 5.2, with a 1 million token context, a 128,000-token maximum output, and continuous adaptive thinking, generally available, and opened Claude Mythos 5.2 to verified organizations. Both models are faster than their predecessors and cost $8 per 1 million input tokens and $40 per output token.

    Usage details

    Who can use it: Fable 5.2 is available on Claude API, Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Foundry; Mythos 5.2 is available only to organizations accepted into Anthropic's verification programs.

    How to use it: Set the `model` field in a Messages API request to `claude-fable-5-2`; for Mythos, apply to the verification program and use `claude-mythos-5-2`.

    Tip: For long tasks, start by measuring with the default `high` effort; move to a higher level only when evals show a quality improvement.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Fable 5.2 and Mythos 5.2Source: Claude Platform release notes
  • Plugin / IntegrationClaude Code cloud sessions

    Memory and reconnection issues fixed in cloud connector and MCP sessions

    Fixes address hangs or out-of-memory errors with connectors that return very long resource names, MCP servers appearing connected without tools after reconnecting, and incorrect reporting of cloud proxy permission status. Large uploads are retried after HTTP 507.

    Usage details

    Who can use it: Claude Code cloud sessions, connectors, MCP servers, and uploads behind a proxy.

    How to use it: Update Claude Code; retest connector resource lists and MCP reconnect flows.

    Tip: Keep resource names short on connector servers and design retries to be idempotent.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • AvailabilityClaude Code cloud sessions

    Messages with large images and subagent messages are preserved after a cloud restart

    If a restart occurs before Claude gives its first response, a message containing a large image is not dropped; a subagent’s pre-restart message does not go unanswered in the main conversation. This improves continuity for cloud sessions after a restart.

    Usage details

    Who can use it: Claude Code cloud sessions and restart/reconnect flows.

    How to use it: After the session reconnects, check the main conversation and subagent history; verify the transcript before resending a task that appears to be missing.

    Tip: Keep large images and critical subagent tasks idempotent to prevent automatic retries from creating duplicate work.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • APIClaude API

    A new toolset is now required for computer use

    Fable 5.2 and Mythos 5.2 support the computer use feature through `computer_toolset_20260801`. The old `computer_20251124` definition returns a 400 error on surfaces other than Amazon Bedrock.

    Usage details

    Who can use it: Claude API, Claude Platform, and supported cloud surfaces; Amazon Bedrock will temporarily continue to accept the old tool.

    How to use it: Define the `computer_toolset_20260801` toolset in requests instead of the old `computer_20251124`.

    Tip: Before moving to production, search all platforms for the old tool name and remove it; don't rely on temporary compatibility in Bedrock.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: What's new in Claude Fable 5.2Source: Claude Platform release notes
  • CapabilityClaude Cowork

    Cowork hooks can no longer change question and plan results

    In Cowork cloud sessions, the PostToolUse hook can no longer change the AskUserQuestion response or ExitPlanMode plan in the result; Claude reads the original result. This prevents hooks from compromising the integrity of user decisions and plans.

    Usage details

    Who can use it: Claude Cowork cloud sessions and organizations using PostToolUse hooks.

    How to use it: Use hooks only for observation or side effects; redesign workflows that rely on changing question and plan results.

    Tip: Read critical approvals from Claude’s original question/plan result, not from hook output.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Pricing / PlanClaude API

    Fable 5.2 and Mythos 5.2 price drops to $8/$40

    The new models' standard price is $8 per MTok of input and $40 per MTok of output, down from the $10/$50 price listed for Fable 5.1 and Mythos 5.1. The Claude Code announcement lists cache reads at $0.20 per MTok.

    Usage details

    Who can use it: Claude API and related Anthropic-managed cloud surfaces; final pricing from cloud providers may vary by provider.

    How to use it: Change the model ID in your existing client to `claude-fable-5-2` or, with authorized access, `claude-mythos-5-2`; monitor usage costs in Console.

    Tip: For long inputs, measure prompt caching alongside the new pricing to compare whether switching from the old model is actually cheaper.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Platform release notesSource: Claude Code changelog 2.1.321
  • FeatureClaude API

    API fast mode enabled for Fable 5.2 and Mythos 5.2

    The output token speed of the two new models can be increased by up to 2.5x with the `speed: "fast"` option in Messages API requests. Fast mode is available for Fable 5.2 and Mythos 5.2 at $16 per MTok of input and $80 per MTok of output.

    Usage details

    Who can use it: Claude API, Anthropic and Google managed Google Cloud; through `-fast` model deployments in Microsoft Foundry. Mythos access is limited to verified organizations.

    How to use it: Add `speed: "fast"` to the Messages API body; in Microsoft Foundry, call the `claude-fable-5-2-fast` or `claude-mythos-5-2-fast` deployment and do not send the `speed` field.

    Tip: Use Fast mode for streaming and latency-sensitive work; note that it is not supported together with the Batch API or Priority Tier.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Platform release notesSource: Fast mode
  • CapabilityClaude Code

    If Fast mode is limited, it falls back to standard speed on the same model

    When Fast serving is rate-limited, overloaded, or in error, Claude Code does not lose the response, switch models, or retry for minutes; it first tries the same model at standard speed.

    Usage details

    Who can use it: Claude Code 2.1.321 and supported models that use Fast mode.

    How to use it: Use Fast mode as usual; no additional setting is required for fallback to work.

    Tip: When measuring speed SLAs, record Fast and standard fallback events as separate metrics.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • SettingClaude apps gateway

    Gateway error messages no longer expose credentials

    Claude apps gateway error messages show the gateway address without a username or password. This preserves useful endpoint information for support while preventing secrets from leaking into error text.

    Usage details

    Who can use it: Claude apps gateway 2.1.321 and Claude Code/Claude Desktop behind the gateway.

    How to use it: Get the gateway address from the error message; check only secure gateway logs for authentication details.

    Tip: Still mask usernames and passwords in URLs in log collection systems; logs from older versions are unaffected.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude apps gateway

    Gateway upstream endpoint is now set with `base_url`

    The gateway no longer considers the environment's `ANTHROPIC_*_BASE_URL` variables for Bedrock, Vertex, and Foundry upstreams. Administrators who want to use a different endpoint must explicitly set `base_url` in the upstream definition.

    Usage details

    Who can use it: Administrators using a Bedrock, Vertex, or Foundry upstream on Claude apps gateway.

    How to use it: Add `base_url` to the gateway upstream configuration and verify that it does not rely on environment variables.

    Tip: For proxy migrations, define the gateway config explicitly in version control rather than silently relying on environment variables.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude apps gateway

    Layered gateway policies combine MCP servers

    Under `policy_mode: layered`, managed MCP servers are now collected across all matching policies, including `match: {}`. A gateway using this mode may expose more managed servers after an upgrade.

    Usage details

    Who can use it: Claude apps gateway and enterprise organizations using layered policy.

    How to use it: Review your gateway policy file; if you use `policy_mode: layered`, test the MCP merge and access boundaries across all matching policies.

    Tip: Export the active MCP list before and after the upgrade to check for unexpected tool expansion.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Managed marketplace names and plugin sources are enforced

    A name defined in managed `extraKnownMarketplaces` cannot be added from another source, hides the entry from the external source, and reinstalls plugins listed by the marketplace from there. Data inherited from another source is also set aside for cleanup; settings and secrets can be restored with the restore command.

    Usage details

    Who can use it: Claude Code 2.1.321 for organizations using managed settings.

    How to use it: Define the `extraKnownMarketplaces` source in the organization setting; use the `claude plugin restore-set-aside` command to restore set-aside data.

    Tip: Keep marketplace names unique and stable across the organization; don’t rely on personal marketplaces with the same name.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    MCP OAuth and Bash permission bypasses closed

    Secrets such as refresh tokens can no longer be sent to another host through MCP OAuth redirects; issues with read-only Bash escaping the sandbox, writes evaluated in the wrong directory, and deny/protected-file checks on cp/mv targets have been fixed. This release also resolves various permission prompt races more accurately.

    Usage details

    Who can use it: Claude Code 2.1.321; MCP OAuth, Bash sandbox, and file permission policies.

    How to use it: Update Claude Code; retest OAuth redirect hosts for MCP servers and Bash deny/ask rules.

    Tip: Add redirect, symlink, and different working-directory scenarios to security regression tests.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • CapabilityClaude API

    Thinking block behavior during model switches clarified

    On Claude API and Google Cloud Google-managed surfaces, reasoning from previously supported models is preserved when a conversation switches to Fable 5.2 or Mythos 5.2. When a conversation switches from these models to another model, new thinking blocks are discarded, but the request succeeds.

    Usage details

    Who can use it: Claude API and Google Cloud Google-managed; transition rules in the model documentation apply to other providers.

    How to use it: Configure your model router to send `claude-fable-5-2`/`claude-mythos-5-2` thinking blocks only to the same two model families.

    Tip: For every route that switches models, validate scenarios that preserve thinking blocks and scenarios that discard them with separate regression tests.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5Source: Claude Platform release notes
  • SettingClaude Code

    Mods activate without approval in auto and bypass modes

    Mods written by Claude now open without asking for approval in auto and bypass modes, even if some managed restrictions such as `allowUnsandboxedCommands: false` or `allowManagedDomainsOnly` are in place. Mods are not loaded in `dontAsk` mode.

    Usage details

    Who can use it: Claude Code 2.1.321; auto, bypass, and dontAsk sessions using mods.

    How to use it: Check permission behavior in `/config` and the permission mode in the organization’s managed settings; if you don’t want mods without approval, use a more controlled mode instead of `dontAsk`.

    Tip: Also test that mods’ own tool calls are subject to deny rules and organization policy in your enterprise policies.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Mods’ own tool calls don’t show permission prompts

    `$.tool.call` and `$.mcp.call` calls in mods do not generate permission prompts under the user’s ask rules, `dontAsk`, or `-p` execution. Deny rules and organization policy still apply.

    Usage details

    Who can use it: Claude Code 2.1.321 and terminal, headless, or SDK sessions using mods.

    How to use it: Define the security boundary with deny rules and managed policy; don’t rely on user prompts alone to protect mod calls.

    Tip: Especially in `-p` and auto mode, restrict the MCP tools mods can call with a narrow allowlist.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Session startup, terminal commands, and git status are more reliable

    Fixes address issues where a failed git status appeared clean or the first file was mistakenly marked as staged, multiple Bash commands ran in the wrong directory, and zsh `KSH_OPTION_PRINT` produced empty output. On Linux NFS failure, `claude -p` now returns an error and 1 instead of silently returning 0.

    Usage details

    Who can use it: Claude Code 2.1.321; sessions using Bash, git, zsh, and Linux NFS.

    How to use it: Update to the new version; handle error codes in `claude -p` text/JSON output, and don't treat git status unavailable as an error.

    Tip: In CI, check that the transcript and git status output were actually produced before treating exit code 0 as success.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • SettingClaude Code

    The `--settings <file>` flag is added for Remote Control

    `claude remote-control` now accepts a settings file to apply to every session it starts. If the file changes, no new session is started; Remote Control must be restarted for the changes to take effect.

    Usage details

    Who can use it: Claude Code 2.1.321 and terminal/remote clients using `claude remote-control`.

    How to use it: Run `claude remote-control --settings /path/to/settings.json`; restart the process when the settings file changes.

    Tip: Keep the shared settings file in version control and don't put secrets in it as plain text.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Remote Control sessions stay in their starting directory

    A connected Remote Control client can no longer move a session to a different directory from the one where it started. This prevents the remote client from operating in the wrong workspace.

    Usage details

    Who can use it: Claude Code 2.1.321 Remote Control sessions.

    How to use it: Start Remote Control from the correct working directory; start a new session to change directories.

    Tip: Record the starting directory as an absolute path in automation commands, and don't try to reuse the session connection.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Remote Control and Desktop sessions recover more reliably after disconnects

    A crashed Remote Control process can receive a new message sooner; Queue for later messages no longer slip into the middle of an active turn; and issues where a disconnected session appeared connected or showed login expired have been fixed. Messages getting stuck in Desktop Code tab and repeated `/compact` calls have also been fixed.

    Usage details

    Who can use it: Claude Code 2.1.321 Remote Control, web/mobile clients, and Desktop Code tab.

    How to use it: Update Claude Code and the Desktop client; check the connection status again in a new session.

    Tip: For long tasks, track connection status separately from session history loading; a connected badge alone isn't enough.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • Pricing / PlanClaude

    Paid plans get temporary Fable usage and limit resets

    From November 9–16, Pro and Team standard seats can use Fable models up to 50 percent of their weekly limits at no additional cost and get one Fable reset. Max, Ultra, Team premium, and legacy seat-based Enterprise premium seats get between one and three resets, depending on subscription duration.

    Usage details

    Who can use it: Pro, Max, Ultra, Team standard/premium, and legacy seat-based Enterprise premium; Free, self-serve Enterprise, usage-based Enterprise, and applicable standard Enterprise seats are excluded.

    How to use it: In Claude, open Settings > Usage and select `Reset for free` under Limit resets.

    Tip: A reset does not raise your limit; it only restores usage spent on Fable. Save it for a genuinely demanding task before it expires.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Fable 5.1, Claude Fable 5Source: Claude Fable 5.2 launch offers
  • AvailabilityClaude

    New models added to the Covered Models scope

    Fable 5.2 is designated a generally available Covered Model, and Mythos 5.2 a limited-access Covered Model. For these models, a default minimum 30-day data retention period and automated safety review apply; zero data retention is unavailable unless Anthropic explicitly authorizes it.

    Usage details

    Who can use it: Fable 5.2 is available in Claude apps, Claude Platform, Bedrock, Google Cloud Agent Platform, and Microsoft Foundry; Mythos 5.2 is limited to approved partners.

    How to use it: Before selecting a model, check Covered Models and your organization’s retention agreement; if you need ZDR, discuss the Enterprise Frontier Safeguards option with your Anthropic account team.

    Tip: For work requiring compliance, choose models based not only on quality but also on retention and safety review requirements.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Covered ModelsSource: Claude Fable 5.2 and Mythos 5.2
  • APIClaude API

    Prefix check for preserved thinking is now required on new models

    Fable 5.2 and Mythos 5.2 validate submitted thinking blocks whose system prompt, tools, or preceding message prefix has changed, regardless of account age, and return a 400 error when they don't match. This extends behavior that was the default only for certain new accounts on Fable 5.1.

    Usage details

    Who can use it: Claude API and surfaces that support the new models; applications that produce thinking blocks and resend them in a subsequent turn are affected.

    How to use it: Do not change the message prefix before switching models; if a change is required, follow the migration guide or consider the supported beta option `thinking.block_binding.prefix_mismatch_behavior: "drop_block"`.

    Tip: Test router or fallback code to ensure it does not automatically carry over old thinking blocks after switching models.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: What's new in Claude Fable 5.2
  • CapabilityClaude Code

    cp and mv invoked by path now use the same security checks

    Using a full path such as `/bin/cp` is now subject to the same path checks as short `cp` and `mv` calls. This reduces the chance of bypassing the permission policy by writing out the full command name.

    Usage details

    Who can use it: Claude Code 2.1.321; Bash commands and file permission rules.

    How to use it: Define both command behavior and target directories in permission rules; do not make separate exceptions for commands invoked by full path.

    Tip: In security tests, try the `cp`, `/bin/cp`, `mv`, and `/bin/mv` variants in the same scenario.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude API

    Advisor tool requires model compatibility with new models

    When using Fable 5.2 or Mythos 5.2 as the executor, the advisor must be one of these two models; any other advisor model returns 400. Fable 5.2 can be used as an advisor for any executor model.

    Usage details

    Who can use it: Claude API applications using the beta advisor tool.

    How to use it: When migrating the executor to the new models, set the `model` field in the advisor definition to `claude-fable-5-2` or `claude-mythos-5-2`.

    Tip: Validate executor and advisor model pairs at application startup to reduce the risk of receiving a 400 in production.

    ✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Platform release notesSource: What's new in Claude Fable 5.2
  • AgentClaude Code

    Pending subagent reports are answered in a single model call

    Auto mode in headless sessions, Desktop, and VS Code now answers simultaneously pending background subagent reports in a single model call, as in the terminal. This reduces unnecessary calls in parallel agent workflows.

    Usage details

    Who can use it: Claude Code 2.1.321; headless, Claude Desktop, and VS Code auto mode.

    How to use it: Start background subagents as usual; pending reports are grouped automatically.

    Tip: As you track the reduction in call count, also measure the size of reports in each call and context usage.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Dropped images are reread less often in sessions with many images

    In sessions with many images that don't fit in a single request, Claude is told which images it has already seen, and rereading unchanged images is limited. This reduces wasted context and image tool calls.

    Usage details

    Who can use it: Claude Code 2.1.321; image-heavy sessions that don't fit in a single request.

    How to use it: Send images with meaningful file names; refer to the output of the first read instead of asking to read the same image again.

    Tip: For large image sets, send changed files separately and avoid reuploading the same file.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Context estimates and WebFetch errors are reported more accurately

    The overcounting of context usage by tool definitions has been fixed; when the model returns an empty response, WebFetch now reports an explicit error instead of sending it to Claude as page content. Prompt suggestion also avoids unnecessary API calls for tool-call responses.

    Usage details

    Who can use it: Claude Code 2.1.321; sessions using long context, WebFetch, and prompt suggestion.

    How to use it: After updating, recheck the `/context` output and the WebFetch scenario where the model returns an empty response.

    Tip: Do not treat empty web responses as valid sources; use a fallback or retry when an explicit error is returned.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • AgentClaude Code

    Stopped workflow status is reported correctly to Claude

    When a user stops or pauses a workflow, Claude no longer says the workflow is still running. The agent's next step is determined by the actual status.

    Usage details

    Who can use it: Claude Code 2.1.321 workflow and cloud session flows.

    How to use it: Stop a workflow and check Claude's status message; resume only work that is actually running.

    Tip: Show stopped, paused, and running as separate statuses in your operations interface.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • SettingClaude Code

    Email check now blocks only an explicitly risky command

    The account email check no longer asks for confirmation except for a shell command that contains the account email address and names User-Agent. This prevents other commands and scheduled runs from being interrupted by unnecessary prompts.

    Usage details

    Who can use it: Claude Code 2.1.321 and sessions that pass the email address to shell commands.

    How to use it: Review commands that contain email addresses; use a secure secret or stdin instead of putting sensitive information on the command line.

    Tip: Treat this change as a narrowed warning rule, not a security approval; separately prevent email leaks.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Accessibility fixes for screen readers, the prompt panel, and file names

    Fixes address screen reader startup delays, the mode turning off when returning from agent view, and leftover characters in y/n questions being treated as yes. File names containing hidden directional characters are rejected; plugin permission requests are also read aloud explicitly as plugin instead of as a hidden character.

    Usage details

    Who can use it: Claude Code 2.1.321; screen reader, agent view, and file submission flows.

    How to use it: Update to the new version; turn screen reader mode back on and test the startup, question, and file submission flows.

    Tip: In accessibility testing, also cover system clock rollbacks and names containing invisible characters.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • FeatureClaude Code

    Fast mode labels show the current model

    The `/fast`, `/config`, and `/model` screens show the current model's name when Fast mode is supported; the wording in the VS Code row that implied only Opus has been removed. This makes Fast access for Fable 5.2 and other supported models clearer.

    Usage details

    Who can use it: Claude Code 2.1.321 and models that support Fast mode.

    How to use it: Open the `/fast` or `/config` screen; enable Fast for a supported model.

    Tip: Prepare internal usage guidelines that show the speed option alongside its cost; Fast mode is not the standard price.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Pricing / PlanClaude apps gateway

    Gateway spend limit reset time shown in local time

    Claude apps gateway spend limit errors now show the reset time in the user's local time zone, for example, `resets 4pm (America/Los_Angeles)`. The message was simplified so it does not reveal gateway setting names.

    Usage details

    Who can use it: Claude Code and Claude Desktop gateway users.

    How to use it: Check the time and time zone in the limit message; don't confuse the UTC or server time in the gateway log with the time shown to the user.

    Tip: Also record the local time zone in enterprise support records to correctly match the same reset event across different regions.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    An unnecessary rejected self-hosted runner now exits with 0

    If the server indicates that the runner isn't needed because the session has ended or paused, or the job has become invalid, the runner now exits with 0 instead of error code 1. This prevents a normal no-op from being treated as a CI failure.

    Usage details

    Who can use it: Claude Code self-hosted runner 2.1.321.

    How to use it: Treat runner exit code 0 as normal termination in CI automation; monitor logs and separate error codes for actual startup failures.

    Tip: Make sure your automation checks the runner session status as well as the exit code.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Headless sessions reload changed skill mods

    Headless Claude Code sessions reload mods when files in `.claude/skills` change. Use `CLAUDE_CODE_PLUGIN_DIR_WATCH=0` to disable unwanted watching.

    Usage details

    Who can use it: Claude Code headless sessions and mods changed under `.claude/skills`.

    How to use it: Use the default behavior; to disable it, set `CLAUDE_CODE_PLUGIN_DIR_WATCH=0` in the runtime environment.

    Tip: If you use automatic reloading, write mod files atomically to prevent a partially written file from being loaded.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Plugin updates no longer re-download for each worktree

    Plugin auto-updates and `/plugin` marketplace updates no longer re-download a plugin for every worktree or folder in the same repository. Dependencies for the updated plugin are also installed.

    Usage details

    Who can use it: Claude Code 2.1.321; projects using marketplace plugins and multiple worktrees.

    How to use it: Update the plugin marketplace as usual; check that the plugin is not re-downloaded and dependencies are installed across worktrees in the same repository.

    Tip: Verify the cache directory and plugin lock behavior after the update using a clean worktree.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • Plugin / IntegrationClaude Code

    Nested skills folders inside plugins are no longer loaded

    Claude no longer loads `.claude/skills` folders inside the plugin directory under `~/.claude/plugins` when working with files at that location. This change prevents plugin resources from being mistakenly treated as project skills.

    Usage details

    Who can use it: Claude Code 2.1.321; plugins that use nested skills under `~/.claude/plugins`.

    How to use it: Move plugin skills to the supported plugin root and check the list of loaded skills with `claude --debug`.

    Tip: When updating plugin packages, regression-test the skill path and visible command menu together.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Only the beginning and end are shown for long background commands

    Background command notifications and TaskStop results exceeding 1,000 characters now send Claude only the beginning and end of the output. This reduces unnecessary context consumption by long logs.

    Usage details

    Who can use it: Claude Code 2.1.321; background commands longer than 1,000 characters.

    How to use it: When needed, read the full output from the log file written by the command; do not use the notification text as a complete log.

    Tip: For automated analysis, save full logs to a file and have Claude read the relevant line range.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code for VS Code

    VS Code shows a warning when two credentials are configured

    When both an auth token and an API key are configured, the VS Code extension displays a warning above the prompt box, matching the one in the terminal. This makes it clear which credential will be used.

    Usage details

    Who can use it: Claude Code for VS Code 2.1.321 and users who configure both a token and an API key.

    How to use it: Check `ANTHROPIC_AUTH_TOKEN` and API key definitions in VS Code settings and environment variables; keep only your chosen authentication method.

    Tip: On enterprise machines, check project, user, and process environments separately; clearing only the VS Code setting may not be enough.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Programs launched on Windows handle PATH settings more safely

    When locating programs such as git launched by Claude Code on Windows, relative and UNC PATH entries and network share links are skipped; PATH changes in settings are also taken into account. This reduces the risk of running the wrong executable and problems with network paths.

    Usage details

    Who can use it: Claude Code 2.1.321 on Windows.

    How to use it: Add program paths to the trusted PATH in settings; do not use relative or UNC PATH entries.

    Tip: On enterprise machines, compare `git --version` and Claude Code's initial command duration before and after the update.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Code

    Background agent that cannot create a worktree now shows the full error

    If a `claude --bg --worktree` session cannot create a worktree, the row in the Agents view now shows the error from the beginning instead of an internal label or the last 200 characters. This makes it easier to diagnose the cause in the UI.

    Usage details

    Who can use it: Claude Code 2.1.321 Agents view and `--bg --worktree`.

    How to use it: Open the failed row in the Agents view; fix the repository or permission settings based on the full error shown at the beginning.

    Tip: In CI reports, save the entire error text, not just the last line.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AgentClaude Code

    Skills, commands, and agents that disappeared after worktree changes return

    Project skills, commands, and agents reappear mid-session when their folder is removed and restored; an issue where they disappeared from the menu after Claude entered a worktree in SDK and Desktop sessions has also been fixed. This affects workflows that use git stash or worktree switching.

    Usage details

    Who can use it: Claude Code 2.1.321; terminal, SDK, and Desktop sessions.

    How to use it: Update Claude Code, restore the folder, and reopen `/agents` or the skill or command menu.

    Tip: In worktree automation, check skill/agent visibility before starting a task; don't rely on an old-version cache.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • Plugin / IntegrationMicrosoft 365 connector

    Teams write permissions can be made permanent in new Microsoft 365 setups

    For Microsoft 365 connectors added for the first time after November 9, an organization administrator can set the Teams send, post, and reply tools to Always allow; members can also select Always allow for these tools if permitted. HIPAA organizations are excluded from this exception.

    Usage details

    Who can use it: Claude Team and Enterprise organizations setting up a new Microsoft 365 work/school connector after November 9, 2026; excludes HIPAA scope.

    How to use it: Go to Organization settings > Connectors > Microsoft 365; set the Teams send/post/reply tools to Always allow, Needs approval, or Blocked.

    Tip: Use Always allow only for low-risk tools with role-based restrictions; leave email and calendar write operations set to Needs approval.

  • Pricing / PlanClaude apps gateway

    Spend is recorded sooner on a busy gateway replica

    The gateway spend limit check skips fewer checks on a busy replica and records spend sooner. This makes the limit message appear closer to the actual usage.

    Usage details

    Who can use it: Organizations using Claude apps gateway and enforcing spend limits.

    How to use it: Update the gateway; monitor the spend view and replica logs together.

    Tip: Set limit alert thresholds with a small margin to account for replica delays, rather than exactly at the limit.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Ultra plan and upgrade text now appear in the Claude Code interface

    The descriptions for session-feedback, `allowedProviders`, and `forceLoginMethod` now include Ultra; Upgrade to Max options have been updated to Upgrade to Max or Ultra. This aligns plan access and sign-in text with the current paid plans.

    Usage details

    Who can use it: Claude Code 2.1.321; relevant organization settings for Pro, Max, Ultra, and other paid plans.

    How to use it: Update Claude Code; plan or sign-in method text updates automatically, and you don't need to enable any additional settings.

    Tip: Update documentation for users so that references to Max alone also include Ultra.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Desktop Code tab

    Desktop Code tab can send up to 20 messages to other sessions

    The Claude Desktop Code tab session increased the number of messages it can send to other sessions before requiring the user to type again from 10 to 20. Long-running parallel workflows require less manual intervention.

    Usage details

    Who can use it: Claude Desktop Code tab and sessions compatible with Claude Code 2.1.321.

    How to use it: Update the desktop app; the behavior is enabled automatically.

    Tip: For teams using many background messages, monitor context usage and waiting behavior.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Claude Desktop in ~/Applications is detected on macOS

    The `/desktop` command no longer acts as if Claude Desktop is missing when it is installed in `~/Applications`. Users are directed to the correct Desktop app.

    Usage details

    Who can use it: Claude Code 2.1.321 on macOS.

    How to use it: Install Claude Desktop in `~/Applications` and run the `/desktop` command in Claude Code.

    Tip: For enterprise deployments, include both `/Applications` and `~/Applications` paths in installation checks.

    ✓ Verified by official sourceSource: Claude Code release v2.1.321
  • FeatureClaude Code

    Older models are grouped under More models in the model picker

    In the `/model` picker, More models is now a row at the bottom that does not renumber the other rows. Older models open under this row.

    Usage details

    Who can use it: Claude Code 2.1.321.

    How to use it: Type `/model` and select the More models row to open the list of older models.

    Tip: Write numbered model-selection instructions in team documentation using model names; row numbers may change between versions.

    ✓ Verified by official sourceSource: Claude Code changelog

Google · Gemini

#2 Gemini 4 Carbon#8 Gemini 4 Argon
Full history (244) →

New in this edition · 6

  • Plugin / IntegrationGoogle AI Studio

    Google Maps Platform connects to Google AI Studio with one click

    Google AI Studio can now set up Google Maps Platform directly in the Build experience and manage keys. Developers can create a live prototype using Maps, Places, Routes, and Environment tools simply by giving Gemini an app idea as a prompt.

    Usage details

    Who can use it: Available to Google AI Studio Build developers, with no credit card required in the trial environment and a quota of 100 calls per API per day. A paid Google Cloud billing account is required for production and higher capacity.

    How to use it: Go to https://ai.studio/build and describe your map-based app idea to Gemini. Enable Google Maps Platform in AI Studio with one click or connect an existing Google Cloud project; then run the live preview.

    Tip: First, prototype using the built-in sandbox quota; when moving to production, enable only the Maps APIs you use and the required billing.

  • CapabilityGemini Enterprise Agent Platform

    Fast mode is now generally available for Anthropic Claude models

    Google has made Fast mode generally available for supported Anthropic Claude models. In online Pay-as-you-go predictions, the "speed": "fast" setting speeds up output generation; the feature initially supports Claude Fable 5.2.

    Usage details

    Who can use it: Available for Claude Fable 5.2 on global, US multi-region (us), and Europe multi-region (eu) endpoints in Gemini Enterprise Agent Platform, for online Pay-as-you-go requests. Provisioned Throughput is not supported.

    How to use it: Enable the Gemini Enterprise Agent Platform API in your Google Cloud project and select Enable for Claude Fable 5.2 in Model Garden. Add "speed": "fast" to the top-level body of the online prediction request and send it to the global, us, or eu location.

    Tip: Use it for time-critical interactive agents; note that Fast mode costs 2x the standard token rate and works only with Pay-as-you-go.

  • ModelGemini Enterprise Agent Platform

    Anthropic's Claude Fable 5.2 is now generally available on Agent Platform

    Claude Fable 5.2 is now GA in Model Garden. The model is designed for technical and scientific work, coding, and finance; it supports text, image, and PDF inputs, with up to 1 million input and 128,000 output tokens.

    Usage details

    Who can use it: GA in Model Garden; available on the global endpoint and the US (us) and Europe (eu) multi-region endpoints. Approval of Anthropic's terms and the Advanced AI Safety Addendum is required on the model card.

    How to use it: Open the Anthropic > Claude Fable 5.2 card in Model Garden; accept the Advanced AI Safety Addendum and relevant terms, then select Enable. Use "claude-fable-5-2" as the model ID in API requests.

    Tip: Use the large context window to review long technical documents, PDFs, or large codebases in a single context; check the model's safety and cost terms before production use.

  • FeatureGemini

    Gemini shopping features are now available in Chinese in Taiwan

    Google has added a Chinese-language shopping experience to the Gemini app in Taiwan. Responses powered by Google Shopping Graph provide product carousels, cross-site price and feature comparisons, availability, reviews, and purchase links.

    Usage details

    Who can use it: Available in Chinese to Gemini app users in Taiwan; no additional requirements based on Google AI plan or platform were specified.

    How to use it: Use the Gemini app in Chinese to write a natural-language shopping query that includes your budget, intended use, and technical criteria. Gemini creates a comparison table and a carousel with product images and links.

    Tip: Include your budget, size, use case, and priorities in the same prompt; prices and availability may change, so check the retailer link again before purchasing.

  • ModelGemini Enterprise Agent Platform

    Meta's Muse models are available in preview on Agent Platform

    Google has made Meta's Muse Wonder 1.0, Muse Ember 1.0, and Muse Spark 1.3 models available as Preview on the global endpoint. All three models support reasoning, function calling, structured output, and a 1M-token context window; Wonder and Spark accept text/image/PDF, while Ember accepts text/image.

    Usage details

    Who can use it: Preview in Gemini Enterprise Agent Platform Model Garden; available only through the stateless OpenAI-compatible Responses API on the global endpoint. Meta models are not Google products and are subject to separate model terms.

    How to use it: Enable the relevant Meta model card in Model Garden. Send a POST request to "https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/meta/v1/responses" and use "muse-wonder-1.0", "muse-ember-1.0", or "muse-spark-1.3" in the model field; stateful "store: true" is not supported.

    Tip: Start with Muse Wonder for complex planning and coding, Muse Ember for high-volume subtasks, and Muse Spark for MCP and long-context agent workflows.

    ✓ Verified by official sourceModels: Muse Wonder 1.0, Muse Ember 1.0, Muse Spark 1.3Source: Gemini Enterprise Agent Platform sürüm notları — 9 Kasım 2026Source: Meta'dan Muse modelleri
  • SettingGemini

    Gemini is rolling out an editable Memory summary setting

    9to5Google reported that Gemini began rolling out the Memory summary feature on November 9; Google's help page confirms that the feature supports viewing, editing, and deleting what Gemini Apps understand about the user, as well as reviewing Memory history. However, the rollout information for that date was not included in official release notes, so the report was not considered verified.

    Usage details

    Who can use it: According to Google's help page, available to users aged 18 and over in the US who have an active Google AI plan, use English, and have Keep Activity and Memory enabled; in Gemini web and mobile apps.

    How to use it: Go to gemini.google.com and follow Settings > Personal Intelligence > Memory summary. From there, you can add or remove details with a prompt and review changes through Memory history; you can also use the "What do you know about me?" prompt.

    Tip: Check Memory history regularly and use a prompt to delete details that are no longer relevant or are sensitive; to turn off Memory, go to Settings > Personal Intelligence.

OpenAI · ChatGPT

#4 GPT-6 Astra#5 GPT-6.1 Sol#7 GPT-6.1 Terra
Full history (339) →

New in this edition · 32

  • AvailabilityChatGPT Enterprise

    ChatGPT Enterprise is now available on AWS Marketplace

    ChatGPT Enterprise can now be discovered and purchased through AWS Marketplace, using existing AWS purchasing and billing relationships. Eligible customers can also access ChatGPT for Healthcare and ChatGPT for Financial Services through the same channel.

    Usage details

    Who can use it: Eligible AWS customers; ChatGPT Enterprise, plus Healthcare and Financial Services products subject to eligibility verification.

    How to use it: Open the ChatGPT Enterprise listing on AWS Marketplace; confirm eligibility and launch steps with your OpenAI or AWS account team, then proceed using your existing AWS purchasing/billing account.

    Tip: Before purchasing, verify together the spending terms that may count toward AWS commitments and your organization’s OpenAI contract.

  • APIOpenAI Go SDK

    Go SDK Agents supports asynchronous function tool calls

    Go SDK v3.98.0 supports configuring asynchronous function tools in the Agents API, where the result is completed later.

    Usage details

    Who can use it: Applications using the Agents API with OpenAI Go SDK v3.98.0.

    How to use it: Upgrade the SDK to v3.98.0; configure the function tool in your Agents session to complete asynchronously, then deliver the relevant tool result to the session when it is ready.

    Tip: When running long-running jobs in a queue, include your own job ID in the tool call and design retries to be idempotent.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • SettingOpenAI Go SDK

    Obfuscation and padding options added to Go SDK Agents streams

    Obfuscation can be configured in Agents event-stream responses, and padding is exposed for 33 event types. The `include_obfuscation=false` query option can be used to disable obfuscation in existing session streams.

    Usage details

    Who can use it: OpenAI Go SDK v3.98.0 and applications using Agents session streams.

    How to use it: Upgrade to v3.98.0; use the stream query with obfuscation enabled by default, or add `option.WithQuery("include_obfuscation", "false")` for backward compatibility.

    Tip: If you parse stream events persistently, validate the event type and actual payload separately for events with padding instead of relying on a fixed schema.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • SettingOpenAI Go SDK

    Prompt cache TTL can now be selected in Go SDK Agents sessions

    v3.98.0 adds the ability to select a prompt-cache TTL of 30 minutes or 12 hours for Agents sessions, and inspect the selected TTL in session and turn responses.

    Usage details

    Who can use it: OpenAI Go SDK v3.98.0 and Go applications using the Agents API.

    How to use it: Upgrade the `github.com/openai/openai-go/v3` package to v3.98.0; select a supported `30m` or `12h` TTL in the Agents session creation or turn settings, then read the active value from the returned session/turn field.

    Tip: Prefer a 12-hour TTL for long, repetitive workflows and 30 minutes for short, variable prompts to avoid unnecessary caching.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • APIOpenAI Go SDK

    Go SDK adds OCI Vault support for external organization keys

    External organization key management can now configure keys backed by Oracle Cloud Infrastructure Vault through the Go SDK. This adds a new provider to the previously documented AWS, GCP, and Azure options.

    Usage details

    Who can use it: OpenAI Go SDK v3.98.0 and eligible API organizations using external organization key management.

    How to use it: Upgrade the SDK to v3.98.0; select the Oracle Cloud Infrastructure Vault backend in the external organization key configuration and specify the relevant Vault credentials and key reference.

    Tip: Grant Vault access permissions only to the required OpenAI organization workloads, and test the key rotation process independently of the OpenAI configuration.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • SettingOpenAI Go SDK

    Noise reduction strength can now be selected in Go SDK Realtime sessions

    An input noise-reduction strength setting has been added for Realtime, transcription, and translation sessions.

    Usage details

    Who can use it: Applications using Realtime, transcription, or translation sessions with OpenAI Go SDK v3.98.0.

    How to use it: Upgrade to v3.98.0; select the input noise-reduction strength value in the relevant session configuration and start the audio session.

    Tip: Because aggressive reduction can distort speech details with clean microphones, measure separate values in quiet and noisy environments.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • CapabilityOpenAI Go SDK

    Image context compression can be configured in Go SDK Responses requests

    Image context compression policies can be configured in Responses create, compact, and WebSocket requests. The active policy is reported in the response.

    Usage details

    Who can use it: OpenAI Go SDK v3.98.0 and Go applications using the Responses API.

    How to use it: Upgrade to v3.98.0; add an image context compression policy to a create, compact, or WebSocket Responses request, then record the active policy field returned.

    Tip: Enable compression to reduce costs and context overflow with large image sequences; for detail-sensitive tasks, test quality while monitoring the active policy.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • APIOpenAI Go SDK

    Go SDK enables author-model and reasoning-effort options for SVG edits

    The author-model and reasoning-effort options can be used for SVG output in Images edit requests.

    Usage details

    Who can use it: OpenAI Go SDK v3.98.0 and Go applications using the image editing API.

    How to use it: Upgrade the SDK to v3.98.0; in an SVG image edit request, select SVG as the output type and send the `author-model` and `reasoning-effort` options as needed.

    Tip: To balance editing accuracy and latency, measure and set reasoning-effort values according to the task type.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • APIOpenAI Python SDK

    Python SDK adds stream options for Agents sessions

    openai-python v3.48.0 adds configuration of stream options for Agents session streaming requests.

    Usage details

    Who can use it: OpenAI Python SDK v3.48.0 and Python applications using Agents session streaming.

    How to use it: Upgrade with `pip install openai==3.48.0`; add the SDK’s new stream options object to the Agents session streaming call and select the desired event-stream behavior.

    Tip: In production, upgrade client and server SDK versions together; safely ignore new event types you do not support.

  • APIOpenAI Go SDK

    Go SDK recognizes `max_output_tokens` turn errors

    The SDK now recognizes errors caused by `max_output_tokens` in Agents turn responses and distinguishes them in the typed error flow.

    Usage details

    Who can use it: Applications using Agents turns with OpenAI Go SDK v3.98.0.

    How to use it: Upgrade the SDK to v3.98.0; when handling turn errors, add a separate branch for the `max_output_tokens` error code and adjust the token limit or replan the task.

    Tip: Instead of retrying this error as a general server error, change the output limit and task-splitting strategy.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • OtherOpenAI Go SDK

    Go SDK fixes mutation behavior for optional request union fields

    In v3.98.0, pointers returned from request unions are bound to the stored optional values, so changes are now reflected in the serialized request. Explicitly provided zero values are preserved, while omitted and null values still return `nil`.

    Usage details

    Who can use it: All API clients using OpenAI Go SDK v3.98.0.

    How to use it: Upgrade the SDK to v3.98.0; change an optional value obtained from a union, then serialize the request. Test an explicit zero value separately from an omitted field.

    Tip: If you reuse request objects, make a copy before each call to verify that mutations do not carry over to the next call.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.98.0 sürüm notları
  • AgentOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds an A2A 1.0 adapter

    Through `@openai/agents/a2a`, native agents can call remote agents with A2A 1.0, serve a native Agent, or integrate deferred remote work. The adapter uses JSON-RPC polling.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2; authorization, persistent storage, and execution are application-specific.

    How to use it: Import the `@openai/agents/a2a` module; set up the A2A client or server adapter, and define the JSON-RPC polling endpoint and your own authorization.

    Tip: Restrict the side effects a remote agent can perform with an explicit application-side permission list.

  • Plugin / IntegrationOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds native hosted tools and enhanced web search

    v0.22.0-alpha.2 adds the native `fileSearch()`, `codeInterpreter()`, and `imageGeneration()` tools, with streamed progress and provider results. `webSearch()` now supports filters, location, tool selection, and citation details.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2 prerelease.

    How to use it: Upgrade all five Agents SDK packages to the same alpha version; add `fileSearch()`, `codeInterpreter()`, or `imageGeneration()` to agent tools and configure `webSearch()` filters.

    Tip: Check provider and model constraints for hosted tools at application startup; display streamed progress separately in the user interface.

  • RetiredOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK removes SDKRunner and switches to a strict write policy

    The experimental `SDKRunner` bridge and `@openai/agents/migration` imports have been removed; migration to the native Agent, Session, and Turn APIs is required. New Environments use `writePolicy: "strict"` by default, and `legacy` must be explicitly selected for the old remote behavior.

    Usage details

    Who can use it: OpenAI Agents SDK v0.22.0-alpha.2 prerelease; native and classic objects/state are kept separate.

    How to use it: Upgrade all five packages together; replace `SDKRunner` and migration imports with the native API or `@openai/agents/classic`, and explicitly select the `writePolicy` value in the Environment.

    Tip: During migration, test strict execution in a new conversation instead of opening old checkpoints directly, and do not mix package versions.

  • AgentOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds turn steering and RunBudget

    Turns can now be controlled with `turn.steerExact()`, attention wait, and typed interruption results. `maxModelSteps` and the optional `RunBudget` limit subsequent requests from the root job.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2.

    How to use it: Use `steerExact()` or `waitForAttention()` on a turn; set `maxModelSteps` and, if needed, `RunBudget` when running the agent.

    Tip: For manual interventions, record the acceptance receipt and stale turn error to prevent the old turn from being applied again.

  • CapabilityOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds an application-owned SessionStore

    Native conversations can now be recovered using an application-owned `SessionStore`; persistent tool-call records and explicit reconciliation of ambiguous results are supported. `fileSessionStore()` provides a local file adapter.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2; the preview is limited to root conversations and application-managed function tools.

    How to use it: Connect `SessionStore` or `fileSessionStore()`; do not automatically replay a conversation before reconciling ambiguous tool results.

    Tip: Use file-based storage for single-process local development, and a custom store with concurrency guarantees for multi-instance production.

  • AgentOpenAI Agents SDK (Python)

    Python Agents SDK adds A2A, tracing, and turn supervision

    Native A2A client and server adapters are now available with the `a2a` extra. TraceContext, workflow tracing, rate-limit snapshots, turn receipts, exact-turn steering, and attention wait controls were also added.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; the `a2a` extra for A2A, and separate `classic` extras for legacy integrations.

    How to use it: Install the `a2a` extra and start the native A2A adapter; enable `TraceContext`, turn receipts, and exact-turn steering options in the workflow.

    Tip: For side-effecting calls to remote agents, persist turn receipts and track them alongside an idempotency key.

  • FeatureOpenAI Agents SDK (Python)

    Python Agents SDK supports Classic and native execution together

    v0.24.0-alpha.7 makes it possible to use preserved Classic workflows alongside native execution through `agents.classic`. New Environment objects default to `write_policy="strict"`, and an explicit compatibility setting is required for the previous behavior.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7 pre-release; native execution on Linux and macOS, with an explicit legacy policy for older remote Windows workflows.

    How to use it: Upgrade with `pip install --pre openai-agents==0.24.0a7`; call legacy workflows through `agents.classic`, and select `write_policy="strict"` in new Environments or explicitly choose `legacy` if compatibility is needed.

    Tip: Before upgrading in production, explicitly define write roots (`writable_roots`) and verify in a test environment that old checkpoints can be opened under the new policy.

  • CapabilityOpenAI Agents SDK (Python)

    Python Agents SDK leaves conversation history to application storage

    `SessionStore`, `SQLiteSessionStore` and `CustomStorage` let applications store root conversation history and automatic Python tool results in their own storage. Ambiguous tool effects can be explicitly reconciled before restarting.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; the initial profile is limited to root conversations and does not cover native tools, Environments, children or portable checkpoints.

    How to use it: Upgrade the SDK to the alpha release; connect `SessionStore` for persistent history, `SQLiteSessionStore` for a single POSIX machine, or `CustomStorage` for your own backend.

    Tip: For tools with side effects, keep an application-level reconciliation record instead of automatically replaying ambiguous results.

  • SettingOpenAI Agents SDK (Python)

    Python Agents SDK changes default reasoning and MCP approval names

    In the bundled model catalog, default reasoning effort for gpt-6-astra, gpt-6.1-sol and gpt-5.6-sol has increased from `low` to `medium`. MCP `approval_mode` values have been renamed from `auto` to `native` and from `approve` to `preapproved`; the old values are rejected.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; applications using the new model metadata and MCP.

    How to use it: If you need the previous reasoning behavior, set `model_settings={"reasoning_effort":"low"}`; use `native` or `preapproved` in MCP configurations.

    Tip: When adopting the alpha, measure response latency and token consumption; specifically search CI tests for the old approval values.

    ✓ Verified by official sourceModels: gpt-6-astra, gpt-6.1-sol, gpt-5.6-solSource: OpenAI Agents SDK Python v0.24.0-alpha.7 sürüm notları
  • Plugin / IntegrationOpenAI Agents SDK (Python)

    Hosted tools and advanced web search come to Python native agents

    Hosted tools `FileSearch`, `CodeInterpreter` and `ImageGeneration` have been added for native agents; Code Interpreter supports caller policies. `webSearch` can now accept domain, location, tool selection and citation details.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; requires an API key provider and a compatible standard Responses model, excluding Responses Lite and Codex login.

    How to use it: Upgrade the SDK; add `FileSearch`, `CodeInterpreter` or `ImageGeneration` to the agent’s tool list, and configure domain/location/tool-choice filters for web search.

    Tip: Limit hosted tool calls to only the tools needed, and store returned citations before showing them to users.

  • Plugin / IntegrationOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds MCP resource and prompt discovery

    Idle Sessions can list and read MCP resources and templates; MCP prompts can also be discovered. This data remains under application ownership and is not automatically added to model context.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2; requires an idle Session and an MCP server.

    How to use it: Use `listMcpResources()`, `listMcpResourceTemplates()`, `readMcpResource()`, `listMcpPrompts()`, and `getMcpPrompt()` on the Session.

    Tip: Apply server, user, and content security checks before passing resource or prompt data to the model.

  • CapabilityOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK adds a native Session voice workflow

    With `VoicePipeline`, `OpenAISpeechProvider`, and `NativeSessionVoiceWorkflow`, you can create workflows that voice verified final text on a native Session. The Realtime example delegates the main work to the native Session.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2.

    How to use it: Import `VoicePipeline`, `OpenAISpeechProvider`, and `NativeSessionVoiceWorkflow` from `@openai/agents/voice` and connect them to a native Session.

    Tip: Start voice generation only after the turn is complete; do not send intermediate tokens directly to the voice engine.

  • SettingOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK can turn tool timeouts into model results

    When `timeoutBehavior: "error_as_result"` is selected in tool callbacks, a timeout can be passed to the model as a failed tool result instead of terminating the turn by default. The optional `timeoutErrorFunction` provides formatting.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2.

    How to use it: Add `timeoutBehavior: "error_as_result"` and, if needed, `timeoutErrorFunction` to the tool callback settings; define in the prompt or tool contract how the model should handle this result.

    Tip: For operations whose side effects may be incomplete, verify the job ID and idempotency status before retrying after a timeout.

  • CapabilityOpenAI Agents SDK (JavaScript/TypeScript)

    JavaScript Agents SDK expands tracing, turn receipts, and context controls

    TraceContext propagation, tool callback trace context, turn receipts, and provider status and usage observations were added. `responseIntegrityReason`, preserving selected model context, and paused NativeGoals updates are also available.

    Usage details

    Who can use it: OpenAI Agents SDK JavaScript/TypeScript v0.22.0-alpha.2.

    How to use it: Carry `TraceContext` through turns and tool callbacks; record observation fields and `responseIntegrityReason`, and explicitly select the context to preserve after compaction.

    Tip: In distributed agent chains, match turn receipts with trace IDs to separate debugging from provider usage data.

  • CapabilityOpenAI Agents SDK (Python)

    Python Agents SDK adds context, compaction, and completion hooks

    With `Agent.transform_context`, `before_compact`, selected-item preservation, and `before_complete`, you can control the context sent to the model, compaction, and whether the suggested response is accepted. `prepare_request` also allows request editing within a limited HTTP profile.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; hook capabilities depend on the provider and native execution profile in use.

    How to use it: Attach `transform_context`, `before_compact`, or `before_complete` callbacks in the agent definition; modify the sampled request with `prepare_request` only in a supported profile.

    Tip: Use `transform_context` to reduce sensitive or unnecessary context without changing persistent history; account for rejected and already-streamed text remaining in history.

  • SettingOpenAI Agents SDK (Python)

    Python Agents SDK adds RunBudget for root workflows

    `RunBudget` can limit root request acceptance, reported tokens or estimated cost before the next request. A single accepted request can exceed the limit, and children are out of scope.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; native agent root runs.

    How to use it: Create a `RunBudget` when running an agent and define a request, token or cost limit; when the budget stops, provide the user with an explicit stop-or-continue decision in your application.

    Tip: Do not treat RunBudget as a hard billing cap; track provider usage separately for actual costs.

  • Plugin / IntegrationOpenAI Agents SDK (Python)

    Python Agents SDK adds MCP resource, prompt and reconnection operations

    Session can now list and read MCP resources and resource templates, discover MCP prompts, and reconnect to a selected HTTP or stdio server. Retrieved data is not automatically added to the model context.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; idle Session and selected HTTP/stdio MCP servers.

    How to use it: On an idle Session, use `list_mcp_resources()`, `list_mcp_resource_templates()`, `read_mcp_resource()`, `list_mcp_prompts()` and `get_mcp_prompt()`.

    Tip: Before sending data read from MCP to the model, validate its source, user and permissions in your application.

  • CapabilityOpenAI Agents SDK (Python)

    Python Agents SDK adds native Session voice streaming with history

    `NativeSessionVoiceWorkflow` can connect a caller-managed text-output Session to a speech pipeline and voice validated final text. Completion review and output validation are retained in the voice workflow.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; requires the `voice` extra; Realtime agents remain in the classic namespace.

    How to use it: Install the `voice` extra and connect a text-output Session to a speech pipeline with `agents.voice.NativeSessionVoiceWorkflow`.

    Tip: Pass only final text that has passed completion review to speech to prevent intermediate drafts from being voiced.

  • SettingOpenAI Agents SDK (Python)

    Python Agents SDK adds tool timeout, buffered Responses and Codex login options

    With `timeout_behavior="error_as_result"`, a timeout in function tools can be converted into a result the model can handle. `NativeProvider(response_mode="buffered")` can wait for the full HTTP response, and `CodexLogin()` can experimentally reuse a saved Codex login.

    Usage details

    Who can use it: Agents SDK v0.24.0-alpha.7; buffered mode requires an API key and standard Responses model; CodexLogin is experimental and limited.

    How to use it: Set `timeout_behavior="error_as_result"` on the tool; select `response_mode="buffered"` on the provider if needed, or explicitly use `NativeProvider(auth=CodexLogin())` instead of an API key.

    Tip: Use buffered mode only on endpoints that do not require streaming; determine per tool whether a timeout result is safe to retry.

Meta · Meta AI

#9 Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 4

  • AvailabilityInstagram

    AI age estimation expands Instagram Teen Account protections globally

    On November 9, 2026, Meta began expanding its AI age-estimation technology globally, bringing millions of Instagram accounts estimated to belong to teens under Teen Account protections—even if they list an adult birth date. Parents in the US will see an in-feed informational video this month; in Australia, the Apple age-range signal is being used as an additional signal.

    Usage details

    Who can use it: Rollout has begun for Instagram accounts worldwide. Expansion on Facebook is ongoing; similar technology will launch on Threads later this year. The parent information in the US will appear this month, while the Apple signal in Australia will be used as an additional age signal.

    How to use it: There is no user-enabled setting; if Meta’s systems estimate that an account belongs to a teen, it is automatically placed under Teen Account protections. Parents in the US can tap the informational video shown in their Instagram feed to open the linked age-assurance explanation.

    Tip: Parents can review the informational video and linked age-assurance explanation to check why Teen Account protections were applied; since rollout is automatic, don’t look for an opt-in beforehand.

  • ModelMeta Model API ve Muse Code

    Muse Ember is out: Meta's smallest and cheapest Muse model ($0,10 / $0,50)

    Meta Superintelligence Labs released Muse Ember, optimized for tool calls, coding subtasks, routing, data extraction, and long-document processing repeatedly performed by agents throughout the day. The model understands text, image, audio, and video input; it offers 1M token context and up to 512K output. It scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5’s score at about one-third the task cost (~$0,14 per task); it is the first custom model trained with Muse Wonder Full.

    Usage details

    Who can use it: Available from November 9 as `muse-ember-1.0` in the Meta Model API and on all Muse Code subscription plans. Pricing (prompts up to 200K tokens): $0,10 per 1M input tokens, $0,004 for cached input, and $0,50 for output; for longer prompts, $0,30 / $0,012 / $1,50. Also available on Google Cloud Model Garden (preview) and Oracle Cloud. Closed weights.

    How to use it: Create an API key at dev.meta.ai; set `base_url` to `https://api.meta.ai/v1` in your existing OpenAI-compatible client and use `model: "muse-ember-1.0"` (Responses API is supported). Install Muse Code with `curl -fsSL https://dev.meta.ai/install.sh | bash` and select `muse-ember-1.0` as the model in subagent definitions.

    Tip: Meta’s recommended pattern: assign planning and oversight to Muse Wonder, and delegate many well-defined subtasks (file porting, bulk document/image labeling) to parallel Muse Ember agents; in Meta’s example, a 10,000-photo job cost $1,33 instead of ~$20 at Wonder pricing.

  • APIMeta VR Glasses

    Android app preparation and a `metavr` CLI workflow open for Meta VR Glasses

    Meta announced that Android developers can begin adapting their existing apps for Meta VR Glasses as of November 9, 2026. Spatial window interfaces and Look and Pinch interactions can be tested before the hardware arrives using the Meta VR Android Studio Plugin, Meta Spatial Simulator, and `metavr` CLI.

    Usage details

    Who can use it: Android developers can begin preparation and simulation today; Meta VR Glasses hardware is expected in spring 2027. The current plugin build mentioned in the article supports Android Studio Quail 2026.1.1–2026.1.4.

    How to use it: Install the Meta VR Android Studio Plugin in Android Studio; add `vrglasses` for `com.oculus.supportedDevices` in `AndroidManifest.xml`. Then run the simulator and app with the commands `metavr init`, `metavr tools install spatialsim`, `metavr ssim start`, `metavr app install path/to/app.apk`, and `metavr app launch package.name`.

    Tip: Since Meta VR devices do not include Google Mobile Services, separate Google sign-in, notification, and Play Billing dependencies into different product flavors; after using the simulator, also test window sizing and the Look and Pinch flow on real hardware.

  • APIThreads API

    Threads adds light, dark, and automatic theme selection for oEmbed embeds

    The Threads API now supports an optional `theme` parameter to create public post or profile embeds in light, dark, or auto mode. This lets external sites and apps choose the appearance of a Threads embed to match their user interface.

    Usage details

    Who can use it: For Threads oEmbed API developers using public Threads post/profile embeds; options are `light`, `dark`, and `auto`.

    How to use it: Send a `GET /oembed?url=<Threads-post-or-profile-URL>&theme=dark` request to the Threads oEmbed endpoint; set `theme` to `light`, `dark`, or `auto`, then embed the returned `html` field on your site.

    Tip: Use `auto` instead of a fixed theme to match the user’s system or site color preference; use `dark` or `light` when a custom design is needed.

Moonshot AI · Kimi

#10 Kimi K3.1 Pro
Full history (320) →

New in this edition · 15

  • SettingKimi Code

    User and project permission rules can now be persisted

    Permission rules scoped to user and trusted local projects are now loaded at session start and written to a persistent file. addRules makes the write operation awaitable; writing to project scope is not supported in remote environments.

    Usage details

    Who can use it: Kimi Code permission system; user scope applies to ~/.kimi-code/config.toml, while project scope applies only to .kimi-code/local.toml in a trusted local workspace. Project scope is not supported in remote environments.

    How to use it: Define a user rule in `~/.kimi-code/config.toml` or a project rule in `.kimi-code/local.toml` in a trusted local workspace, or use `scope = "user"` or `scope = "project"` when calling `addRules` and await the Promise; start a new session if you changed the file externally.

    Tip: Use project rules only in trusted local workspaces, and check the environment type first rather than assuming a persistent deny rule will be written in a remote environment.

  • CapabilityKimi Code

    Primary directory changes in multi-directory projects now apply to existing sessions

    When the primary working directory changes in a multi-directory project, the existing local session follows the new primary; FS, Git, AGENTS.md, session metadata, and the trust view are updated consistently. Explicit environmentCwd and remote connections retain their fixed-directory behavior.

    Usage details

    Who can use it: Kimi Code multi-directory local projects; shared Core flows across Desktop/Web/CLI. Sessions using explicit environmentCwd or remote connections are not excluded and retain their previous behavior.

    How to use it: In a multi-directory project, change which project directory is primary, then run `/reload` or continue in the existing session; use explicit environmentCwd to pin a specific directory.

    Tip: After changing the primary, check paths relative to the new primary rather than the previous one; pin tasks that need to run from a different root with explicit environmentCwd.

  • FeatureKimi Code

    Same-name Skill variants can now be selected in multi-directory projects

    Skill items with the same name across multiple working directories are now retained as variants in the catalog; TUI, Agent, REST, and SDK can select the right variant with an optional path or workspace_dir. If no path is provided, the existing default priority is preserved.

    Usage details

    Who can use it: Kimi Code TUI, Agent, Web, REST, and SDK clients using multiple working directories; merged into the main branch on November 9.

    How to use it: In a multi-directory project, enter /skill:<name> in TUI and select the listed directory variant; if using REST or SDK, add the optional path and workspace_dir fields to the activation request.

    Tip: When selecting among skills with the same name, check the displayed directory label; existing calls without a path continue to use the default variant.

  • SettingKimi Code

    config.toml permission rules are applied when an Agent starts

    deny, ask, and allow permission.rules defined in config.toml are now actually applied when a new or restored Agent is created. The fix also covers flows created through print mode, TUI, subagent, fork, and /reload.

    Usage details

    Who can use it: Kimi Code CLI’s TUI and print mode, as well as Desktop, Web, and ACP sessions that use the same Core, including subagents that use the rules file.

    How to use it: Add `[[permission.rules]]`, `decision`, and `pattern` fields to `~/.kimi-code/config.toml`; start a new session or run `/reload`, then verify the deny/ask/allow behavior for a matching tool call.

    Tip: First test with narrowly scoped deny rules for sensitive files; note that an already-open session needs to be recreated or reloaded with /reload to see an external config change.

  • CapabilityKimi Code CLI

    `--skills-dir` directories now refresh automatically in open sessions

    Explicit Skill directories passed with --skills-dir in the CLI are now watched automatically in open sessions. New or changed SKILL.md files and the targets of watched symbolic links are reloaded; manual reload is no longer needed in most cases.

    Usage details

    Who can use it: Local macOS and Windows 11 workspaces using --skills-dir with Kimi Code CLI/Core; applies when the existing watch behavior is enabled.

    How to use it: Start Kimi Code with `kimi --skills-dir /path/to/skills`; while the session is running, add a new Skill folder or SKILL.md to the directory, or change a file.

    Tip: If you use symbolic links or Windows junctions, do not disable Skill watching so that both the link itself and changes to its target can be monitored.

  • CapabilityKimi Code

    Images from MCP tools are reused in subsequent requests

    For providers that support it, images returned by MCP tools are now referenced by the file ID from the initial upload in each model request instead of being resent as base64. This reduces request payload size and lowers the risk of timeouts on slow networks.

    Usage details

    Who can use it: Kimi Code CLI and other Core clients using MCP tools that return images; the provider must support image uploads and referencing.

    How to use it: Run an MCP tool that returns an image, such as a screenshot tool, and start a new turn in the same conversation; with a supported provider, the file reference is reused after the initial upload.

    Tip: For repeated screenshot or chart turns, use a provider that supports file references instead of sending the image inline each time.

  • CapabilityKimi Code

    Skill calls are now routed directly through the Skill tool

    Kimi Code’s skill-loading instructions have changed: the model is now directed to call the Skill tool instead of reading the SKILL.md file directly with Read. This ensures contexts such as $ARGUMENTS and ${KIMI_SESSION_ID} are handled through the official activation path.

    Usage details

    Who can use it: Kimi Code CLI, Desktop, Web, VS Code, ACP, and SDK sessions; the change merged into the main branch on November 9 and depends on a subsequent package/canary release.

    How to use it: Use /skill:<name> in Kimi Code; to run a skill, have it call the Skill tool instead of opening the SKILL.md file with Read.

    Tip: Read SKILL.md only for inspection or editing, so parameters and session context within the skill are not lost; use the Skill path to run it.

  • CapabilityKimi Code CLI

    Fullscreen copying now reaches the actual clipboard on Linux and tmux

    Text selected with the mouse and /copy in the fullscreen CLI now use the shared application clipboard path; a false success report on Linux and tmux has been fixed. OSC52 fallback is preserved in remote sessions, and copy results are reported more accurately.

    Usage details

    Who can use it: Kimi Code CLI fullscreen TUI mode; especially Linux, tmux, WSL, Termux, and remote sessions.

    How to use it: Use `tui_mode = "fullscreen"` in tui.toml, select text in the transcript or run `/copy`; if needed, force OSC52 in a remote environment with `KIMI_CODE_CLIPBOARD_OSC52=1`.

    Tip: If you use tmux, also check the outer terminal’s `set-clipboard` or passthrough setting; both OSC52 formats are now sent together.

  • FeatureKimi Code CLI

    Fullscreen transcript groups are pinned and live progress has been added

    Fullscreen transcript groups remain pinned as the stream runs; tool, thinking, and response progress appears below the latest group. The Activity area now also shows elapsed time and the completion time of the finished turn.

    Usage details

    Who can use it: Kimi Code CLI fullscreen TUI; the Activity duration summary has also been improved in both view modes.

    How to use it: Run Kimi Code with the fullscreen TUI and start an Agent task; watch the progress line below the latest transcript group and the duration summary when the turn completes.

    Tip: Use the default grouped view to keep the layout organized during parallel tool calls; increase the Tool call density setting when you need more detail.

  • SettingKimi Code CLI

    The `copy_on_select` setting adds automatic copying in fullscreen mode

    The behavior of copying text immediately when selected in the fullscreen CLI can now be enabled or disabled with copy_on_select. It is enabled by default; when disabled, you can copy the selection with Ctrl+C, Ctrl+Shift+C, or Cmd+C.

    Usage details

    Who can use it: Only the Kimi Code CLI fullscreen TUI; available through `/settings` and tui.toml.

    How to use it: In fullscreen mode, go to `/settings` → Copy on select and choose On or Off; alternatively, add `copy_on_select = false` to tui.toml and run `/reload`.

    Tip: To avoid copying accidental selections to the clipboard in long transcripts, turn off the setting and copy only with an intentional Ctrl+C.

  • AgentKimi Code

    Artifact worker that can write intermediate outputs added to Tower

    Tower can now distinguish artifact worker tasks that are considered read-only in terms of their business purpose but can write intermediate outputs. This brings intermediate artifact generation into the Tower flow without triggering unnecessary human review for every new artifact.

    Usage details

    Who can use it: Kimi Code Tower tasks and Agent flows using Tower workers; PR merged into the main branch on November 9, with availability depending on the next version release.

    How to use it: Start a task that generates artifacts in the Tower environment and use the artifact worker flow as the worker type; track the result in the Tower task’s artifact outputs.

    Tip: Assign outputs that do not require human approval, such as intermediate renders or reports, to an artifact worker to reduce unnecessary pauses in the review queue.

  • SettingKimi Code CLI

    Transcript tool density and thinking-folding settings added

    Under [transcript], options are now available for tool_call_density: compact, balanced, detailed, and ungrouped; and for thinking: grouped and ungrouped. The fullscreen /settings screen shows a sample transcript preview of the options.

    Usage details

    Who can use it: Kimi Code CLI fullscreen TUI; settings apply through config.toml and fullscreen `/settings`.

    How to use it: Open Fullscreen `/settings` → Tool call density or Thinking folding; for persistent settings, use values such as `tool_call_density = "balanced"` and `thinking = "grouped"` under `[transcript]`.

    Tip: Choose compact for narrow terminals and balanced or detailed when inspecting tool details; ungrouped is the fallback for returning to the previous detailed flow.

  • CapabilityKimi Code

    Skill changes in `extra_skill_dirs` now refresh live

    extra_skill_dirs defined in config.toml are now watched in open sessions. New or changed skills in these directories are scanned automatically; when the setting changes, old watchers are removed and the new directories are watched.

    Usage details

    Who can use it: Kimi Code users with extra_skill_dirs in config.toml; applies in CLI/Core sessions when the existing watch key is enabled.

    How to use it: Define additional directories with extra_skill_dirs in config.toml; add or change SKILL.md files in these directories while a session is running.

    Tip: When changing the directory list, verify that old paths are removed; the system now watches only the current extra_skill_dirs sources.

  • CapabilityKimi Code

    Default TowerStatus output for LLMs reduced to the latest 20 records

    When Tower has many workers, the output Agent gets from the TowerStatus tool now returns the latest 20 records by default; the scope can be adjusted with query parameters. Full results are preserved for TowerStatus calls run by humans.

    Usage details

    Who can use it: Agent/LLM flows using TowerStatus; the direct TowerStatus view for human operators is unchanged.

    How to use it: Have Agent use the TowerStatus tool; when a broader or different view is needed, provide the call’s query parameters using the fields specified in the tool schema.

    Tip: In Tower environments with hundreds of workers, start by reviewing the latest 20 records and expand only the required range with a parameter.

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 2

  • Pricing / PlanGrok Bot

    Grok Bot weekly usage limits reset for all users

    The official Grok Bot account announced on November 9 that it had reset the weekly usage limits for all Grok Bot users. The post does not specify the new limit amount, a permanent quota increase, or the next reset time.

    Usage details

    Who can use it: All Grok Bot users; weekly usage pool.

    How to use it: No action is required from users; open Grok Bot and check the Weekly usage limit value under Settings → Usage & Billing.

    Tip: Plan intensive tasks for the start of the new week, and check additional usage or the limit policy before running a large job.

  • CapabilityGrok Bot

    Token efficiency and smarter behavior fixes are rolling out for Grok Bot

    Elon Musk and Matt Palmer announced that fixes to improve Grok Bot’s token efficiency and behavior are rolling out. Musk said the fixes would be applied overnight and that a noticeable improvement should be expected on November 9; no new quota amount or metrics were shared.

    Usage details

    Who can use it: Grok Bot users; plan, country, and permanent rollout scope were not specified.

    How to use it: No activation is required; use Grok Bot as usual. If the issue persists, reply to Musk’s announcement post or Matt Palmer’s feedback post.

    Tip: For long tasks, first compare token usage and completion quality using short, repeatable tasks; note any changes on the quota screen.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 19

  • Plugin / IntegrationAuK

    VoxCanvas content-editing pipeline for AuK released

    The VoxCanvas content-editing pipeline is now available. Refined Emilia transcripts rewritten with Qwen3-ASR-1.7B and containing alignment timestamps can be used for TTS training or the pipeline’s initial edit-annotation stage.

    Usage details

    Who can use it: AuK developers and researchers can use the VoxCanvas repository and the zkniu/VoxCanvas dataset on Hugging Face.

    How to use it: Open the VoxCanvas link in the AuK README, download the dataset, and follow the VoxCanvas pipeline’s TTS training or edit-annotation stage 1 workflow.

    Tip: Before generating edit labels, verify the transcripts’ alignment timestamps against the audio segments; misalignment reduces TTS training quality.

  • Capabilityteamai-cli

    teamai-cli keeps pending records visible during concurrent compaction

    JSONL sidecars remain visible even while they are being read during compaction, preventing new records from being lost or added twice in a race condition. A separate option remains available for calls that specifically want to exclude pending records.

    Usage details

    Who can use it: The teamai-cli main branch dated 2026-11-10; setups that use recall and report records concurrently.

    How to use it: Update teamai-cli to a commit after this change and run your existing recall/report flow as usual; pending records are automatically merged on reads by default.

    Tip: For teams that perform concurrent imports or compaction, don’t manually merge JSONL files after updating; use the CLI’s locking and pending-record flow.

  • ModelHyra-results

    HyPDE released as a 323-million-parameter 2D PDE model

    HyPDE has been added to the Hyra-results model releases. The model predicts the next four frames from the last six field frames on a regular 2D grid and supports autoregressive rollouts for 13 physical systems.

    Usage details

    Who can use it: Researchers and developers; the Apache-2.0-licensed Hyra-results repository, with Python 3.10+, PyTorch 2.4+, and preferably a CUDA GPU.

    How to use it: Run `git clone https://github.com/Tencent-Hunyuan/Hyra-results.git && cd Hyra-results/AI4Science/hypde`; then run `bash weights/download.sh`, `pip install -e .`, and `hypde forecast <file>.hdf5 --index 0 --t0 17 --steps 60 -o forecast.npz`.

    Tip: For long forecasts, track VRMSE against the dataset rather than treating the model output as reliable in one go; no claim is made that it outperforms on all 13 systems.

  • Plugin / IntegrationLoopForge

    LoopForge adds six agent CLIs to chat and SOPs

    LoopForge’s native sandbox driver selects seatbelt on macOS and bwrap on Linux. AI chat and SOP runs can now use OpenCode, Cursor, Antigravity, Kimi, Pi, and DeepSeek Harness with personal credentials; after the first response, the chat stays in the selected CLI.

    Usage details

    Who can use it: The LoopForge platform; users with relevant personal CLI/API keys on macOS and Linux native sandbox installations.

    How to use it: Install the latest LoopForge; open the CLI selector in the AI chat composer, choose OpenCode/Cursor/Antigravity/Kimi/Pi/DeepSeek Harness, and save the relevant key in Personal Settings. Select the same runner for the SOP role as well.

    Tip: Choose the CLI prefix in the model identity from the start rather than changing runners within the same conversation; this avoids the issue where another CLI cannot continue a native session.

  • Capabilityteamai-cli

    teamai benchmarks show pinned judges, calibration, and trial costs

    Each benchmark arm is pinned to the same judge tool and model per scorer; the plan and result record this choice. Before `teamai benchmark run` starts, the estimated cost is shown; results show cost per trial, scorer success rate, and 95 percent intervals, while the dashboard shows judge and calibration details.

    Usage details

    Who can use it: teamai teams using benchmark suites; comparisons run on a local or sandbox driver.

    How to use it: Define `scorers` in suite.yaml, use `check` or `successCriteria` in tasks, and run `teamai benchmark run <suite> --sandbox`; view the results on the dashboard’s Benchmarks page.

    Tip: Keep the same judge settings for every arm when comparing tools/models; otherwise, observed success differences may be due more to the evaluation tool than the production tool.

  • Featureteamai-cli

    teamai delivers resources as folders with their supporting files

    Workflow, hook, rule, and MCP server resources can now be folders containing their own supporting files instead of single files. This lets teams push/pull shared resources alongside additional instructions and references—for example, alongside WORKFLOW.yaml or HOOK.yaml.

    Usage details

    Who can use it: teamai-cli team repositories; teams sharing workflow, hook, rule, or MCP resources with multiple files.

    How to use it: Organize the relevant resource as a folder in the team repository; use workflows/<name>/WORKFLOW.yaml for workflows, hooks/<name>/HOOK.yaml for hooks, rules/<name>/RULE.md for rules, and the relevant folder structure for MCP, then run `teamai push`; have members run `teamai pull`.

    Tip: Keep the folder name the same as the existing resource name; when changing the filename or folder layout, also check the delivery record for the old single-file resource.

  • Retiredteamai-cli

    teamai score --run --queue removed; use driver none

    The `teamai score --run --queue` option has been removed. To start a queue independently of the driver, use `--driver none`; direct calls to `teamai score` from within an agent session are also rejected.

    Usage details

    Who can use it: Teams using teamai score and scorer automation; legacy command-line integrations.

    How to use it: Replace the old `teamai score --run --queue` call with `teamai score --run --driver none`; instead of scoring from within an agent, run it in your own terminal or use teamai-share-session.

    Tip: Check the command’s exit code in CI scripts; the unrecognized `--queue` option now produces an error instead of being silently queued.

  • Capabilityteamai-cli

    Benchmark and workflow tools start with PWD set to their own workspace

    A tool launched by teamai now uses the clone or worktree directory from which it was run as `PWD`. This ensures that OpenCode benchmark/workflow steps, in particular, modify their own isolated copy rather than a member’s checkout.

    Usage details

    Who can use it: Teams using teamai benchmark and workflow; OpenCode or other tools sensitive to the working directory.

    How to use it: After updating, run `teamai benchmark run <suite> --sandbox` or the relevant workflow as usual; check the tool step’s `PWD` value and confirm that written files are in the arm clone.

    Tip: Use `PWD` or relative paths instead of binding workflow commands to fixed absolute directories; this prevents writing to the wrong checkout across arms.

  • AvailabilityFlashTopK

    FlashTopK adds support for NVIDIA SM100/SM103 and Blackwell GPUs

    Scan, pruning, and refine kernels have been added for SM100; FlashTopK can now run on SM100/SM103 hardware with FP8 or BF16. B200 gets tuned support, while GB200, B300, and GB300 fall back to the B200 cost model; SM100 requires a CUDA 12.8+ driver.

    Usage details

    Who can use it: NVIDIA Blackwell SM100/SM103; B200, GB200, B300, and GB300 systems. Applies to supported models using a 32×128 indexer, such as Hy4-preview.

    How to use it: Install FlashTopK from the latest source with `pip install -e .`, use a CUDA 12.8+ driver on an SM100 machine, and check hardware/operand support with `flashtopk.supported(q, kv)`.

    Tip: On SM100 devices other than B200, verify that the device is using the fallback cost model before the first benchmark; adjust the cost manually with `FLASHTOPK_PRUNE_COSTS` if needed.

  • Settingteamai-cli

    teamai hooks no longer generate votes or reminders when Recall is off

    When Recall is off, no adoption record, vote, vote push, or TodoWrite reminder is generated. The disablement order follows `TEAMAI_RECALL_DISABLED=1`, the member setting, and the team setting; `recall enable/disable` takes effect on the next hook event.

    Usage details

    Who can use it: Teams using teamai recall; installations where it is off by default or disabled at the machine/team level.

    How to use it: Temporarily set `TEAMAI_RECALL_DISABLED=1` or use `recallEnabled: false` in `~/.teamai/config.yaml`; for the team level, turn off `sharing.recall.enabled`.

    Tip: After turning Recall off or on, don’t expect the old hook result in the same session; trigger the next prompt or tool event to see the new behavior.

  • APIteamai-cli

    teamai upgrades benchmark plan and run records to version 2

    Benchmark plan and run records are now processed as version 2; runs started by older teamai are not resumed. Although the dashboard can display older records, regression comparisons are limited to runs within the same version.

    Usage details

    Who can use it: Existing setups using teamai benchmark; older benchmark runs in progress after an upgrade.

    How to use it: After upgrading teamai, start a new run with `teamai benchmark run <suite>` instead of resuming the old one, and treat the first result as the new baseline record.

    Tip: Don’t upgrade versions in the middle of a benchmark comparison; the first run on the new version may not have a comparable previous run.

  • Featureteamai-cli

    teamai can generate benchmark suite and task drafts from past sessions

    The new `teamai-sessions-to-tasks` workflow converts previously completed sessions into benchmark task drafts. `teamai session list --finished --all --suite ...` lists local and team sessions; selected sessions can be added to a suite with their starting and solution commits and change size.

    Usage details

    Who can use it: teamai teams creating benchmark suites; completed sessions published to the team repository.

    How to use it: Run `teamai session list --finished --all --suite benchmarks/<suite> --sizes "S=40,3 M=200,10 L=800,30"`; select suitable sessions and draft task files with the default `teamai-sessions-to-tasks` workflow.

    Tip: Before adding drafts to the suite, check that the starting and solution commits are actually from the same repository and that the task correctly identifies protected test files.

  • Capabilityteamai-cli

    teamai reduces hook wait time from about 180 ms to 30 ms

    On macOS and Linux, tool-call and prompt hooks that produce no output run in the background without loading the CLI package. In measurements, the tool-call median dropped from about 177 ms to 32 ms, and the prompt median from about 200 ms to 31 ms.

    Usage details

    Who can use it: macOS and Linux teamai installations; hooks that sync over HTTP or produce output continue to use the normal full flow.

    How to use it: Update teamai-cli and run a tool-call/prompt inside an agent; verify the behavior with `[hook-dispatch] ... fast path` entries in `~/.teamai/debug.log`.

    Tip: Make sure your custom hooks don’t produce output on every tool-call; when output is needed, teamai is expected to use the full wait path instead of its background shortcut.

  • Capabilityteamai-cli

    teamai preserves local files that replace deleted skills

    When the team removes a skill folder using a tombstone, a local file that a member placed in or changed to replace that folder is no longer deleted. `pull` reports the preserved path and the reason.

    Usage details

    Who can use it: Teams that deliver skills/rules with teamai pull; members with locally edited files when a team resource is removed.

    How to use it: After removing the team resource and running `teamai push`, have members run `teamai pull`; review the `Kept ...` message in the output and manually archive or rename the local file.

    Tip: Before removal, run `teamai status` and capture version control state; the preserved file remains in the workspace even though it is no longer managed by the team.

  • Settingteamai-cli

    TodoWrite reminder moved to wildcard PostToolUse instead of a separate hook

    A separate hook entry is no longer installed for TodoWrite; the reminder now runs within `Hook dispatch post-tool-use wildcard`. The old TodoWrite timeout override is ineffective; teams that want to set a timeout must configure the wildcard key.

    Usage details

    Who can use it: Teams using teamai hook configuration; especially the built-in TodoWrite reminder in Cursor, Codex, and WorkBuddy.

    How to use it: Remove the old `Hook dispatch post-tool-use TodoWrite` timeout override from `hooks/hooks.yaml` or set a timeout for `Hook dispatch post-tool-use wildcard`; to disable it, use the TodoWrite key in the builtin disabled list.

    Tip: After updating your custom timeouts, run `teamai doctor`; it explicitly reports the obsolete override.

  • Capabilityteamai-cli

    teamai no longer delivers folders with reserved marker names as resources

    Folders containing reserved marker names such as `RULE.md` or `WORKFLOW.yaml` are no longer delivered under an incorrect resource name; `pull` can skip the folder and remove the copy it previously delivered. Users must use a regular filename to preserve the file.

    Usage details

    Who can use it: Team repositories that distribute rules and workflows with teamai pull.

    How to use it: For example, use `rules/onboarding/overview.md` instead of `rules/onboarding/RULE.md`; likewise, use the documented layout for the actual `WORKFLOW.yaml` in a workflow folder, then run `teamai pull`.

    Tip: When creating folder-based resources, don’t use the reserved marker file directly as the resource content; use a regular name to prevent the resource from being named incorrectly.

  • APIteamai-cli

    teamai doctor fixes incorrect failure code for Codex config errors

    When Codex returns a configuration error for the `hooks/list` working directory, `teamai doctor` can now report it as a note and return exit code 0. Previously, CI checks could fail with exit code 1 even when hooks were reliable.

    Usage details

    Who can use it: teamai installations that use Codex hook validation in CI or maintenance scripts.

    How to use it: After updating teamai, run `teamai doctor`; review the Codex configuration error as a warning/note, and assess trust based on the report contents rather than exit code 0/1.

    Tip: In CI, don’t rely only on the general exit code; also log the Codex config error line from `teamai doctor` output.

  • APIteamai-cli

    teamai strictly validates the recall maintenance threshold value

    `recall maintenance --prune --threshold` now accepts only finite numeric values. Inputs such as `abc`, `NaN`, `Infinity`, and similar values are rejected before maintenance starts, with exit code 1 and a descriptive error.

    Usage details

    Who can use it: All platforms using `teamai recall maintenance`.

    How to use it: Use a finite value, such as `teamai recall maintenance --prune --threshold 0.2`; validate invalid inputs with CLI validation before running them in CI.

    Tip: If you pass the threshold value from the command line, validate the shell variable first to ensure it is numeric and finite; otherwise, maintenance won’t start.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 26

  • SettingQwen Code GitHub Channel

    API budget reserve setting added for GitHub PR monitoring

    GitHub Channel tracks the remaining REST budget in successful API responses and temporarily pauses background PR monitoring when the budget reaches the reserve level. The new `prWatchBudgetReserve` setting defaults to 200; a value of `0` disables this threshold, and monitoring resumes when the budget recovers.

    Usage details

    Who can use it: Qwen Code GitHub Channel; available as an optional channel setting in the Web Shell Bound Sessions panel.

    How to use it: In Web Shell, open the channel editor and enter the reserve in the `prWatchBudgetReserve` field in the Bound Sessions panel; set it to `0` to disable the threshold, then reconnect or restart the channel after changing it.

    Tip: If you use a shared classic PAT, set a higher reserve; 200 is a reasonable safeguard for channels using a dedicated token.

  • FeatureAlibaba Cloud DevOps Intelligent Code Review

    Review Memory for Intelligent Code Review released

    Review Memory stores team rules previously explained in reviews as repository context and reuses them in subsequent reviews. If administrator approval is enabled, candidate memories can be approved before taking effect; memories can be edited, disabled, or deleted.

    Usage details

    Who can use it: Repositories using Alibaba Cloud DevOps Intelligent Code Review; repository administrator settings are required, and no plan or region is specified.

    How to use it: Open Repository Settings > AI Assistant Settings > Intelligent Code Review > Review Memory; enable the feature and, if desired, also enable the memory approval option. Suggest a rule in an AI Assistant MR comment using `Please remember:`.

    Tip: Write a memory only as a lasting team rule with clear scope, rationale, and warning conditions; do not save one-off exceptions.

  • FeatureAlibaba Cloud DevOps AI Assistant

    Requirement clarification and content refinement for work items arrive

    AI-powered Requirement Clarification clarifies missing rules through multi-turn questions, while AI-powered Content Refinement turns verified information into a structured work item. Results are not written without user approval, and uncertain points are retained as Pending Confirmation.

    Usage details

    Who can use it: Alibaba Cloud DevOps work items; project managers can configure AI rules and type-specific clarification templates.

    How to use it: On a work item details page, select AI-powered Requirement Clarification or AI-powered Content Refinement; use Project Settings > AI Tool Configuration for team rules and the relevant work item type settings for templates.

    Tip: If rules are missing, use Requirement Clarification first; if there is sufficient information, go directly to Content Refinement. Be sure to approve the preview comparison before writing.

  • FeatureQwen Code Managed Agent

    Model and context statistics arrive in Managed Agent sessions

    Session diagnostics now show model statistics and request context using persistent, bounded statistics. The Context and Model statistics tabs let you inspect the latest main request context, model usage, and related metrics.

    Usage details

    Who can use it: Qwen Code Managed Agent Web Shell diagnostics and the Managed Agent API that provides them.

    How to use it: Open the Diagnostics window in a Managed Agent session; inspect context usage and model statistics in the Context and Model statistics tabs.

    Tip: When investigating context growth, start with Context statistics; compare the Model statistics tab to review model costs or distribution.

  • FeatureQwen Code Managed Agent

    Managed Agent execution diagnostics show the Activity tree

    Managed Agent now persistently records tool results, file read/write/edit operations, Skills usage, and subcalls on a per-session basis. The Web Shell Activity tab lets you inspect the execution flow through an expandable invocation tree.

    Usage details

    Who can use it: Qwen Code Managed Agent Web Shell and the corresponding session activity API.

    How to use it: Open the Diagnostics > Activity tab in a Managed Agent session; expand an invocation row to inspect tool results, file operations, and subcall details.

    Tip: When reproducing a task, compare the failed invocation first, then the Skills and child invocation branches in the same tree.

  • Plugin / IntegrationQwen Code Managed Agent

    Mem0-based personal memory comes to Managed Agent as an opt-in feature

    Managed Agent sessions can now connect to a separate personal Memory Store and use the remember, recall, get, forget, and search operations. The feature is disabled by default; content is kept user-specific in explicitly enabled sessions, and recall results are marked as untrusted external context rather than instructions.

    Usage details

    Who can use it: Qwen Code Managed Agent; `qwen.managed-agent.memory.enabled` defaults to `false`, and a configured Mem0-compatible memory endpoint and Session context are also required.

    How to use it: Set `qwen.managed-agent.memory.enabled=true` in the administrator configuration and specify a Mem0-compatible endpoint; opt in to a personal Memory Store when creating a new Managed Agent Session.

    Tip: Treat recalled memory only as context; do not store secret keys or directly executable instructions in memory.

  • SettingQwen Code GitHub Channel

    PR comments are delivered to all bound sessions that did not write them

    New conversation and inline review comments on a bound PR are delivered specifically to every bound session that did not actually write the comment. The feature is enabled by default; it can be disabled with `prCommentWatch: false`, and comments written by the session itself are filtered out using the transcript.

    Usage details

    Who can use it: Bound sessions with `deliverToBoundSessions` enabled in GitHub Channel; comment delivery is enabled by default.

    How to use it: Leave `deliverToBoundSessions: true` in the channel configuration; set `prCommentWatch: false` to disable it. After binding a PR, receive new conversation or inline comments as private bound events.

    Tip: Keep this setting enabled if you run review and fix sessions with the same GitHub account; disable it only when needed to prevent comment loops.

  • APIQwen Code CLI

    Qwen Code CLI 0.27.1 released

    Qwen Code CLI version 0.27.1 has been released, bringing together new Channel, Web Shell, Managed Agent, ACP, and CLI capabilities. The release notes state that there are no known breaking changes.

    Usage details

    Who can use it: Qwen Code CLI users; the release became available as an official GitHub release on November 9, 2026.

    How to use it: Upgrade your Qwen Code installation to version 0.27.1 and verify your version with `qwen --version`; use the update command available in your package manager.

    Tip: Before upgrading in production, back up your Channel and Managed Agent settings and test the update in a test workspace first.

    ✓ Verified by official sourceSource: Qwen Code v0.27.1 resmi sürümü
  • Plugin / IntegrationQwen Code Managed Agent

    Archive-based self-service installation for Skills and MCP is here

    Managed Agent can install digest-verified Skill archives and MCP configurations into a Session with a single approval through a self-service flow that is disabled by default. New API surfaces have been added for preparing and importing ZIP/tar.gz files, MCP form fields, and encrypted Workspace settings.

    Usage details

    Who can use it: Managed Agent; the feature is disabled by default with `qwen.managed-agent.extensions.self-service-enabled` and is intended for authorized Workspace/Session administrators.

    How to use it: Enable the feature with `qwen.managed-agent.extensions.self-service-enabled=true`; prepare the archive and review its digest, fill out the MCP form, and complete installation with a single approval in the Session.

    Tip: Pin the digest before installation and import only trusted MCP endpoints; keep the setting disabled by default in production.

  • CapabilityAlibaba Cloud Simple Log Service (SLS)

    SLS automatically indexes new log fields

    Field index auto-discovery automatically detects newly written top-level log fields, indexes them as text fields, and enables statistics. This eliminates the need to configure indexes manually or reindex for every new field; numeric conversion in SQL can be performed at runtime.

    Usage details

    Who can use it: The feature is being gradually rolled out to accounts in SLS Console, and a whitelist may be required; the SDK method is available in all regions without a whitelist. The Python SDK example requires version 0.9.52 or later.

    How to use it: In the target LogStore’s index configuration page in Console, enable full-text indexing, select Auto-discover unspecified fields, and save. In the SDK, use `line.auto_key_detect=true` or Python `auto_key_detect=True`.

    Tip: Since automatically discovered fields are text types, consider safe conversions such as `try_cast` in numeric SQL expressions.

  • Plugin / IntegrationAlibaba Cloud EMR Serverless Spark

    Support for multiple Catalogs in a single Spark SQL query arrives

    EMR Serverless Spark’s multi-Catalog capability provides access to DLF, Hive Metastore, MaxCompute, Hologres, StarRocks, Paimon, Iceberg, and custom Catalogs within a single Spark Session. Users can read from and join data across different systems in the same SQL statement, then write to another Catalog.

    Usage details

    Who can use it: EMR Serverless Spark workspaces; supported Catalog types and connection details are managed through Data Catalog.

    How to use it: In the workspace Data Catalog, select the Catalog type and enter the connection details; use tables in the form `<catalog>.<database>.<table>` in SQL to run read, join, and write operations within the same Spark Session.

    Tip: Keep connection and credential details in the centralized Data Catalog and use only logical three-part table names in queries; this lets you preserve SQL when the backend changes.

  • APIQwen Code TypeScript SDK

    TypeScript SDK v0.1.31 released, packaging CLI 0.27.1

    Version v0.1.31 of the Qwen Code TypeScript SDK has been released. This package includes Qwen Code CLI 0.27.1, built from the same branch/ref as the source.

    Usage details

    Who can use it: Developers using the Qwen Code TypeScript SDK; the official SDK release was published on November 9, 2026.

    How to use it: Upgrade your TypeScript SDK dependency to v0.1.31; verify in the release metadata that the bundled CLI version is 0.27.1.

    Tip: Reduce the risk of protocol incompatibilities by using the CLI version bundled with this release instead of pinning the SDK and CLI to separate versions.

  • FeatureQwen Code Web Shell

    Bulk pin and archive selection arrives in the sidebar

    The Web Shell session sidebar now lets you select multiple sessions and pin or archive them in bulk. Existing single-row deletion and search behavior is unchanged.

    Usage details

    Who can use it: Qwen Code Web Shell session sidebar.

    How to use it: Enable multi-select mode in the sidebar, select sessions, and choose the bulk Pin or Archive action.

    Tip: Select only completed sessions before archiving; pin ongoing sessions to keep them visible.

  • CapabilityQwen Code OpenTUI

    OpenTUI executes prompts submitted by Mods

    The OpenTUI terminal now queues `prompt.submit` prompts submitted by a Mod and runs them as separate turns when the terminal is idle. Pending prompts are limited to eight turns, and continuation notification lines are displayed.

    Usage details

    Who can use it: Qwen Code’s experimental OpenTUI terminal and `prompt.submit` Mod hooks.

    How to use it: Enable OpenTUI and the Mod that uses `prompt.submit`; prompts submitted by the Mod are automatically queued when the active turn ends and executed as a new turn.

    Tip: To prevent a Mod from generating too many automatic prompts, submit a small number of clearly scoped prompts per task.

  • AgentPoS (Progression of States)

    PoS detects belief trapping and enables recovery on long-running tasks

    PoS, introduced by Alibaba Cloud and academic partners, monitors belief trapping—when agents continue acting without making progress—using an explicit belief state, Belief Health, and Trapping-Aware Recovery. The project code was released as open source; the paper’s experiments used four benchmarks and three foundation models.

    Usage details

    Who can use it: Open-source PoS project and research implementation; no commercial plan or hosted Alibaba Cloud service is specified.

    How to use it: Clone the PoS GitHub project linked in the official announcement and follow the setup and reproduction steps in the README; in your own agent loop, implement the validated state, unresolved gap, and recovery constraint flows as separate layers.

    Tip: Measure not just repeated tool calls, but whether changes in validated beliefs contribute to the goal; otherwise, loops using different tools may go unnoticed.

    ✓ Verified by official sourceModels: Qwen3.7-Plus, Kimi-K3, GLM-5.3Source: Alibaba Cloud PoS belief trapping duyurusuSource: PoS açık kaynak kod deposu
  • AvailabilityQwen Code Desktop

    Qwen Code Desktop v0.27.1 released

    Qwen Code Desktop version 0.27.1 was released using the same release commit as CLI 0.27.1. The Desktop package brings the user-facing Qwen Code features in this release to the desktop application.

    Usage details

    Who can use it: Qwen Code Desktop users; the official Desktop v0.27.1 release was published on November 9, 2026.

    How to use it: Download the v0.27.1 asset for your operating system from the Desktop release page, or update your existing Desktop installation to this version.

    Tip: Back up your local settings and session history before updating; test installations that use Channel or MCP in particular.

  • FeatureQwen Code Web Shell

    Workspace is now visible in scheduled tasks

    In single-workspace or host-locked workspace configurations, Scheduled Tasks cards and new/edit forms now display the workspace label. A host lock is also marked as Fixed by host; the field is read-only when workspace selection is unavailable.

    Usage details

    Who can use it: Qwen Code Web Shell Scheduled Tasks; single-workspace, host-locked, and multi-workspace views.

    How to use it: Open the Scheduled Tasks page and check the Workspace label on the task card; in the New or Edit form, hover over the label to see the read-only workspace information and full path in single-workspace or host-locked configurations.

    Tip: Check the Workspace and Fixed by host information before editing a task to avoid running it in an unexpected workspace.

  • FeatureQwen Code Managed Agent

    View Managed Agent session resources and child-agent traces

    A diagnostic surface has been added to show Session Resources and direct child-agent traces for Managed Agent. The Resources and Child agents tabs in the Web Shell session diagnostics window let you inspect persistent discovery history and child-agent executions.

    Usage details

    Who can use it: Qwen Code Managed Agent and Web Shell diagnostic surface; sessions you have permission to access.

    How to use it: Open a Managed Agent session, open the diagnostics/Diagnostics window, and switch to the Resources or Child agents tab; API consumers can use the relevant session diagnostics GET surfaces.

    Tip: If a skill or MCP server behaves unexpectedly, compare the Resources history first, then the relevant Child agents trace.

  • Plugin / IntegrationQwen Code Web Shell

    Scheduled tasks now support selecting a channel delivery target

    An optional channel and user/chat delivery target have been added to the Web Shell scheduled task creation and editing flow. Delivery can be disabled; the selected target is retained on task cards, and the feature works in persistent mode.

    Usage details

    Who can use it: Users of Scheduled Tasks in Qwen Code Web Shell; channel delivery and target selection are available in the supported persistent operating mode.

    How to use it: In Scheduled Tasks, open the new task or edit form, select the Channel field and a recently used or manually entered user/chat target; leave Delivery disabled to turn off delivery.

    Tip: Verify the target ID with a small test task before sharing the task; check that the channel and target are retained when you reopen the card.

  • APIQwen Code ACP bridge

    ACP replay responses now include projection watermark information

    An optional `replayProjectionWatermark` observation has been added to full in-memory replay payloads, and the TypeScript client preserves this field. Consumers can now distinguish clean history with the same snapshot and cursor from history that includes a summary projection.

    Usage details

    Who can use it: ACP bridge and TypeScript clients that consume this field; the field is present only in eligible full replay responses.

    How to use it: Read the `replayProjectionWatermark` field in the ACP replay payload response; treat an absent field differently from a value of `0`, and consider only events beyond the watermark as reliable additional history.

    Tip: Use the watermark as a projection boundary, not as proof that the history is complete; PR does not explicitly implement history merging.

  • SettingQwen Code

    Default background execution for subagents is now configurable

    A user setting is now available to determine whether subagents run in the background when this is not explicitly specified in their definitions. The default behavior is selected with the `agents.defaultBackground` value; an explicit run or frontmatter value still takes precedence.

    Usage details

    Who can use it: Agent and subagent workflows using Qwen Code 0.27.1; the setting enables background execution by default.

    How to use it: Add `agents.defaultBackground: true` or `false` to your settings file and restart Qwen Code; an explicit value for a single run or in frontmatter overrides this setting.

    Tip: Use `true` for long-running independent tasks and `false` for short subtasks whose results need to be passed to the parent turn immediately.

  • FeatureQwen Code Web Shell

    A Preview action has been added to renderable artifact cards

    Renderable artifact cards, such as HTML, Markdown, and images, can now be previewed with a clearly labeled Preview action and an eye icon. The details prompt also follows the Web Shell interface language more consistently; external URL cards that cannot be previewed remain Open.

    Usage details

    Who can use it: Renderable artifact results in Qwen Code Web Shell; the existing Open behavior is retained for unsupported types.

    How to use it: When an artifact card appears, click the Preview button or eye icon on the card; use the Ask for details action if an explanation is needed.

    Tip: Open long HTML output with Preview first instead of asking for it again in chat; don’t expect Preview for external URL artifacts.

  • SettingQwen Code CLI

    LSP can be enabled or disabled with a user setting and environment variable

    Experimental LSP support can now be managed with the user-scoped `tools.lspEnabled` setting and equivalent environment variables. It is disabled by default; safe mode still disables LSP, and changes require a restart.

    Usage details

    Who can use it: Qwen Code CLI 0.27.1; CLI users who want to use the LSP feature.

    How to use it: Change `tools.lspEnabled` in your settings or use `QWEN_CODE_ENABLE_LSP=1`; to disable it, set `QWEN_CODE_DISABLE_LSP=1` and restart the CLI.

    Tip: Start by testing with a single project and a trusted LSP server; when troubleshooting, you can temporarily disable LSP with `QWEN_CODE_SAFE_MODE=1`.

  • APIQwen Code daemon ve Web Shell

    Sessions can be sorted by the last question

    The daemon session API now supports the optional `lastUserPromptAt` field and sorting by `updatedAt` or `lastUserPromptAt`. The Web Shell offers Recent activity and Last question options, storing the preference per daemon.

    Usage details

    Who can use it: Qwen Code daemon, TypeScript client, and Web Shell session list.

    How to use it: Select Recent activity or Last question from the sort selector in the Web Shell session list; SDK/daemon consumers can use the `sortBy` value in a list/catalog request.

    Tip: For long-running tasks that update automatically, Recent activity is more useful; use Last question to prioritize the most recent request written by a person.

  • CapabilityQwen Code Channels

    Sessions bound to PRs receive built-in guidance based on their role

    Sessions bound to a GitHub PR with the `review`, `verify`, or `fix` role receive built-in guidance appropriate to their role, even if the operator has not defined an action template. Event summaries now report facts instead of giving the session direct instructions to make corrections.

    Usage details

    Who can use it: Sessions bound to a PR with a role in Qwen Code GitHub Channel; tool permissions are unchanged.

    How to use it: Bind a session to a GitHub PR with a role; the role and its guidance will appear in the session prompt even if you don’t define an action template. Existing templates take precedence over built-in guidance.

    Tip: Don’t give write tools to sessions with the `verify` role; built-in guidance provides behavioral assistance, not an authorization boundary.

  • FeatureQwen Code Web Shell

    Collapse and drag to reorder Web Shell workspaces

    Controls to collapse/expand workspace headers in bulk and reorder them by dragging have been added to the Web Shell sidebar. The order is saved in daemon storage; the collapsed state is specific to the browser.

    Usage details

    Who can use it: Qwen Code Web Shell workspace sidebar; ordering persists across restarts, while the collapsed state persists in the same browser context.

    How to use it: In the Web Shell sidebar, use the workspace header’s toggle control; drag the header to the desired position or choose the bulk collapse/expand action.

    Tip: Move frequently used workspaces to the top and collapse the others to keep the session list shorter on mobile or narrow screens.

Watchlist

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; ranked 11th with 44 points. Available in preview on Google Agent Platform as `muse-wonder-1.0`.

Claude Haiku 5.5

36
Anthropic · Claude

Anthropic’s fast, affordable model at $0.10 / $0.50. Meta says the new Muse Ember delivers the same score at about one-third the task cost.

Muse Ember

36
Meta · Meta AI

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

Grok 4.7

35
xAI · Grok

The Grok app’s default model; text outputs on the Grok API receive a default watermark (`x-watermark`). Grok Bot’s weekly limits were reset for everyone on November 9.

Tencent Hy4

35
Tencent · Hy

Available via API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, free on WorkBuddy and CodeBuddy through November 15.

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba’s flagship cloud model; Qwen 4 has not launched yet. Qwen Code 0.27.1 added optional Mem0 personal memory to Managed Agent.

Sources