YZ·RADARyz.mikail.net
Daily edition · 18 November 2026

GPT-6.2 Astra enters second place with a score of 62, while GPT-6.2 Sol also makes the TOP 10

After a quiet week, the TOP 10 is moving again. OpenAI’s new models entered the upper ranks of the Artificial Analysis Intelligence Index v5.0.1 directly: GPT-6.2 Astra is second with a score of 62, and GPT-6.2 Sol is fifth with 57. Artificial Analysis no longer shows the older GPT-6 Astra (57) and GPT-6.1 Sol (55) in the current list; the new versions have taken their places. As a result, both have dropped out of the TOP 10. Claude Fable 5.2 remains at the top with 65 points. Gemini 4 Carbon (60) fell to third place, and Claude Opus 5.5 (59) to fourth; neither score changed. Claude Sonnet 5.5 (55), Gemini 4 Hydrogen (54), GPT-6.1 Terra (51), Gemini 4 Argon, and Muse Wonder Full (50) held their positions from sixth to tenth. The watchlist is unchanged: Kimi K3.1 Pro is in 11th place with 46 points and remains the strongest open-weight model, followed by Muse Wonder at 44, Qwen4 Max at 43, Claude Haiku 5.5 and Muse Ember at 36, and Grok 4.7 at 35. LMArena’s live page was again blocked by CAPTCHA; the latest accessible view was dated November 11.

GPT-6.2 Astra is positioned for difficult, lengthy tasks, while GPT-6.2 Sol is aimed at everyday work and agent-scale use. Both are available in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans; in Enterprise and Edu, an administrator must enable the model. In the API, Astra costs $8 / $40, less than GPT-6 Astra, while Sol keeps its old price at $2 / $10. Both models have a 1.05M context window and 128K output. For inputs over 272K tokens, input and cached-input prices double, while output costs 1.5 times as much. According to Artificial Analysis, Astra costs $4.00 per task and Sol $1.01. Astra for Law has also moved to the GPT-6.2 Astra base and is now available for any Enterprise customer to request through the Trusted Access Program. OpenAI SDKs (Python 3.53, Node 7.58, Go 3.101, Ruby 0.126) added discounted 2- and 8-hour `completion_window` options for Responses background requests. Agents SDK Python 0.25.0 contains breaking changes: the Runloop sandbox provider has been removed and the checkpoint schema has moved to 1.22; older 0.24 workers cannot read these checkpoints. The Dots administrator preview ended on November 18, and Dots was enabled by default in eligible Enterprise, Edu, and Healthcare workspaces. Standard, Advanced, and Expert dots cost $50, $200, and $500 per month. One OpenAI help page says the free period lasts until January 13, 2027, while another says billing began on November 18; confirm with your account team before signing a contract. In addition, ChatGPT for Sales launched with more than 30 partner integrations, the academic researcher program is expanding to 200,000 people (applications open November 23), and Codex 0.170.0 added WebRTC voice calls in the TUI and recovery from Cloud connection drops.

Anthropic has launched linked billing in beta on Claude Platform. A third-party platform can bill a customer’s API usage to the customer’s own Claude organization without sharing an API key; the customer can set a spending limit per platform and remove the connection from Console. Claude Desktop now has separate Stable and Latest update channels. Claude Code v2.1.327 accepts configuration from stdin with `claude mcp add-json`, lists broken configuration files in `/status`, and asks for confirmation before `/export` overwrites a file. On Google’s side, `service_tier=deferred` is now generally available in the Interactions API: queued calls with a 24-hour completion target are 50% cheaper. The same option is also GA for agents built with google-adk v2.10.0 and later. In BigQuery, `optimization_mode => MINIMIZE_COST`—which reduces costs for AI.IF and AI.CLASSIFY—as well as AI.SEARCH reranking and AI.RANK are now generally available; the Model Garden allowlist requirement has been removed for AlphaGenome. Meta launched Muse in Australia on iOS, Android, web, and Mac, and expanded the Canva connector globally. Instagram introduced the Content Protection tool, which automatically finds copies of original videos for eligible creators.

xAI’s 45 entries are from Grok Bot v1.2.0. Team administrators can turn off proactive messages and manage network mode and blocked sites from the Security section. Enterprise connectors now connect with SSO, and command approval shows the full command and an “Allow once” option. On Tencent’s TokenHub, the built-in `web_search` tool for DeepSeek Vendor Direct models has been charged against Token Plan credits since November 18; failed, empty, and timed-out calls are free. `minimax-m2.7` was removed on November 17, with `minimax-m3` recommended as its replacement. LoopForge v0.4.0 added run history for scheduled tasks, an MR node for code-delivery SOPs, and an approval step for external MCP write operations. AI-Infra-Guard also added new scanning rules for MCP Inspector (CVE-2025-49596) and one-api (CVE-2026-81027). Alibaba’s Qwen4-Max arrived on OpenRouter (`qwen/qwen4-max`), Novita, You.com, Runware, and other providers; its enterprise pilot program also opened. Moonshot released Kimi Code 2.6.3, DeepSeek released Harness v0.2.2-rc.4, and Z.ai published ZCode 3.18.2 packages on GitHub. Xiaomi had no new user-facing changes between November 16 and 18. The Tencent agent returned invalid output on its first run, so it was run again and its complete results were added. A total of 167 new developments were added to the archive today.

Today’s TOP 10

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

  1. 01

    Claude Fable 5.2

    Anthropic · Claude=
    65pts

    Launched on November 9 and leads the rankings with 65 points. Available on all paid Claude plans; `claude-fable-5-2` in the API, $8 / $40, with thinking always adaptive. On November 12, the Claude Meetings beta (Google Meet first) opened to Team and Enterprise plans.Released: 9 November 2026

  2. 02

    GPT-6.2 Astra

    OpenAI · ChatGPTNEW
    62pts

    Newly launched and moved straight into second place with 62 points, replacing GPT-6 Astra (57). At Max settings, it costs $4.00 per task. Available in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans; `gpt-6.2-astra` in the API, $8 / $40, cached input $0.40. 1.05M context, 128K output; reasoning effort `none` is not accepted. The Pro 500 plan includes Ultrafast mode.Released: 17 November 2026

  3. 03

    Gemini 4 Carbon

    Google · Gemini▼ 1
    60pts

    Its score is unchanged (60), but it fell to third place with the arrival of GPT-6.2 Astra. `gemini-4-carbon` is available in Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, $4 / $20 starting December 1). Carbon handles planning and difficult decisions in Antigravity Orchestrator; Orchestrator (Carbon) is available on Ultra.Released: 28 October 2026

  4. 04

    Claude Opus 5.5

    Anthropic · Claude▼ 1
    59pts

    Its score is unchanged (59), but it fell to fourth place. Anthropic’s flagship Opus model, available on Claude Pro/Max/Team/Enterprise and through the API. Claude Code 2.1.322 added fast mode support on Vertex (global, US, EU).Released: 22 September 2026

  5. 05

    GPT-6.2 Sol

    OpenAI · ChatGPTNEW
    57pts

    Newly launched and entered fifth place with 57 points, replacing GPT-6.1 Sol (55). At Max settings, it costs $1.01 per task, one-quarter the price of Astra. `gpt-6.2-sol` in the API, priced the same as GPT-6.1 Sol: $2 / $10, cached input $0.10. 1.05M context, 128K output. Available in ChatGPT Work and Codex; the Pro 200 plan includes Ultrafast mode for Sol.Released: 17 November 2026

  6. 06

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

    The strongest model available on Claude.ai’s free plan as well; sixth place with 55 points.Released: 28 September 2026

  7. 07

    Gemini 4 Hydrogen

    Google · Gemini=
    54pts

    Launched on November 12; scores 54 points with the xhigh setting and costs $3.16 per task ($5.48 for Carbon). `gemini-4-hydrogen` is also available on the free API tier, with paid pricing of $1.50 / $7.50. The top model in the Gemini app for Google AI Plus subscribers. Since November 16, it has also been selectable in Gemini Enterprise AlphaEvolve experiments; it is not the default model and must be explicitly selected in Generation settings.Released: 12 November 2026

  8. 08

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

    Released on October 26 for ChatGPT Work, Codex, and the API; focused on vision and computer use. There is no GPT-6.2 version yet. Dots was enabled by default in eligible Enterprise, Edu, and Healthcare workspaces on November 18.Released: 26 October 2026

  9. 09

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available on Google AI Pro/Ultra and the paid API; promotional price of $2 / $10 through November 30. In Antigravity, Orchestrator (Argon) is available to both Pro and Ultra subscribers.Released: 30 September 2026

  10. 10

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta’s most capable Muse model; no public API. This week, the Muse app added the ability to call businesses for paid subscribers in the US and the Muse from Meta Chrome extension.Released: 12 October 2026

Key signals

  • GPT-6.2 Astra (62) entered second place, and GPT-6.2 Sol (57) fifth; GPT-6 Astra and GPT-6.1 Sol dropped off the list.
  • Gemini 4 Carbon (60) fell to third place, and Claude Opus 5.5 (59) to fourth; Claude Fable 5.2 (65) remains at the top.
  • OpenAI: Dots is enabled by default in eligible Enterprise, Edu, and Healthcare workspaces; ChatGPT for Sales launched.
  • Google: Deferred tier is generally available in the Interactions API and ADK agents, with a 50% discount.
  • Tencent: DeepSeek web search became a paid feature on TokenHub, MiniMax-M2.7 was removed, and LoopForge v0.4.0 launched.
  • Meta enabled Content Protection on Instagram. 167 new developments: xAI 45, Tencent 29, Anthropic 28, OpenAI 28, Alibaba 14, Google 9, Meta 6, Moonshot 5, DeepSeek 2, Z.ai 1.

Worth trying today

01

Keep costs under control when moving to GPT-6.2

In the API, change the model field from `gpt-6.1-sol` to `gpt-6.2-sol`; because the price is the same ($2 / $10), a direct switch is safe for most workflows. Try `gpt-6.2-astra` for the most difficult tasks, but remember that its per-task cost is about four times Sol’s. Don’t set `reasoning.effort` to `none` in requests; the API rejects this value. Try to keep inputs below 272K tokens; above this threshold, all requests become more expensive. Put fixed system prompts and tool definitions at the beginning of the input; cached input costs $0.10 with Sol. Send non-urgent batch jobs using `background: true` and `background_options.completion_window` with the discounted 2- or 8-hour window; you’ll need to upgrade your SDK (Python 3.53, Node 7.58, or later). In ChatGPT Work and Codex, simply select GPT-6.2 Sol or Astra from the model picker.

OpenAI · ChatGPT
02

Run non-urgent Gemini calls at half price with Deferred tier

For jobs that can wait, such as overnight reports, batch classification, or in-depth research, send requests to the Interactions API global endpoint: add `service_tier: "deferred"`, `background: true`, and `store: true` to the body of `POST /v1/projects/PROJECT_ID/locations/global/interactions`. Use the returned interaction ID to query the result, receive it as a stream, or cancel it. The target completion time is 24 hours, status is retained for 7 days, and the price is half that of Standard PayGo. If you’re building agents with ADK, upgrade google-adk to v2.10.0 or later and set `ServiceTier.DEFERRED` in RunConfig. Keep flows that need to respond to users instantly on the Standard tier.

Google · Gemini
03

Review your web search and MiniMax settings on Tencent TokenHub

If you use DeepSeek Vendor Direct models with Token Plan, each successful Search and Fetch call now consumes credits. Remove the `web_search` parameter from requests that don’t need search, cache results where appropriate, and monitor the number of successful searches in your usage dashboard. Since `minimax-m2.7` was removed on November 17, change the model value to `minimax-m3` in your code and configuration. Don’t rely on automatic migration; add retries for brief timeouts during the transition, and verify M3’s output and price against a few examples before sending production traffic.

Tencent · Hy
04

Protect your Instagram videos with Content Protection

In Instagram, go to Home > Tools > Your professional tools > Content protection. If the menu isn’t visible, apply at instagram.com/apply-content-protect. After joining, your new original videos are protected automatically; you can also select eligible older videos separately. The tool finds visual matches on Instagram and Facebook. For each match, choose whether to track it, hide it, add a link crediting you, or release it. If you have older videos you think were copied, flag those first.

Meta · Meta AI

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

New in this edition · 28

  • Pricing / PlanClaude Platform

    Linked billing launches in beta on Claude Platform

    Third-party platforms can bill customers’ API usage directly to their Claude organizations without sharing API keys. The customer’s organization admin approves the connection, can set a spending limit per platform, and can remove the connection from Console.

    Usage details

    Who can use it: Claude Platform beta; available on supported third-party platforms and Claude Platform organizations.

    How to use it: The customer starts the Claude connection flow on a supported platform, links their Claude Platform organization, and sets a monthly spending limit. The developer creates an OAuth app in Console and implements the flow in the linked-billing documentation and the `user-profiles-billing-2026-10-01` beta header.

    Tip: Link the organization instead of giving an API key to a third party, and set a low initial spending limit for each platform.

  • SettingClaude Code

    `/export` prompts before overwriting an existing file

    `/export` now asks for confirmation before overwriting an existing file. An export with a conflicting custom name preserves the file and indicates that you must run `/export` for a confirmed export.

    Usage details

    Who can use it: Claude Code 2.1.327 terminal.

    How to use it: Type `/export` to export the transcript, then confirm or cancel at the existing-file warning. To use a custom name, use `/export <name>`.

    Tip: Use descriptive, date-stamped filenames to avoid overwriting important transcripts.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • FeatureClaude Code

    `/status` lists malformed configurations and disabled components

    `/status` now shows malformed agent, skill, command, and output style files; keybinding errors; an unavailable sandbox; ignored permission settings; and hook matchers that never match, all in one list.

    Usage details

    Who can use it: Claude Code 2.1.327 terminal sessions.

    How to use it: Type `/status` in the Claude Code terminal and fix the file, permission, or hook warnings in the diagnostics section.

    Tip: If a skill or hook isn’t working, checking the `/status` list first is faster than trial and error.

  • APIClaude Code

    `claude -p` proceeds after 10 seconds if model settings get stuck behind a proxy

    If fetching model settings on the first request gets stuck for a reason such as a proxy, `claude -p` no longer waits indefinitely; it continues without those settings after 10 seconds.

    Usage details

    Who can use it: `claude -p` runs launched headlessly, via the SDK, by cron, or by the desktop app.

    How to use it: Run `claude -p` as usual; if the model settings endpoint is unavailable, the request is sent after a 10-second wait.

    Tip: Because this behavior can hide a proxy failure, monitor network and model settings logs separately in production automation.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    The stage at which failed hooks stop is now visible

    Failures from hooks such as SubagentStop, SessionEnd, CwdChanged, and FileChanged are no longer silent in the relevant flows. Failed hooks, hooks that stop Claude, and normal hook results are shown with different colors and descriptions.

    Usage details

    Who can use it: Claude Code terminal and fullscreen view; sessions that use hooks.

    How to use it: Run a hook and check the colored status lines in the transcript or the Ctrl+O view; fix the cause using the hook output.

    Tip: Run hooks in a small test session first, then address the first cause shown in the red error line.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • APIClaude Code

    Billing group naming comes to Claude apps gateway sessions

    With the Claude apps gateway’s `admin.billing_group_tag` setting, a session’s billing group can be named using `attributionProject`. This makes it easier to break down usage by project or cost center.

    Usage details

    Who can use it: Claude apps gateway deployments using `admin.billing_group_tag`.

    How to use it: Enable the billing-group configuration expected by the gateway, then set `attributionProject` in the Claude Code session settings to a project or cost center name.

    Tip: Keep values consistent with your organization’s existing project or charge-code names to make it easier to consolidate reporting.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • AvailabilityClaude Desktop

    Stable and Latest release channels come to Claude Desktop

    Claude Desktop now lets you choose an update channel. Stable gets more thoroughly tested releases less frequently, while Latest delivers fixes and features earlier with roughly daily releases.

    Usage details

    Who can use it: Available on macOS and Windows on all plans; not available on Linux, Microsoft Store installations, or computers managed by organization policy.

    How to use it: In Claude Desktop, open `Settings > Desktop app > Release channel` and select `Latest` or `Stable`. To check immediately, select `Claude > Check for Updates` on macOS or `Help > Check for Updates` on Windows.

    Tip: Choose Stable for daily use, and Latest to try new features early and report issues.

  • SettingClaude Code

    GitHub App installation asks for confirmation before changing a repository

    `/install-github-app` now displays an explicit confirmation before making changes to a repository. This lets you review the target before the app runs repository, secret, or API key steps.

    Usage details

    Who can use it: Claude Code 2.1.327; sessions using the GitHub App installation flow.

    How to use it: Run `/install-github-app` in Claude Code; review the repository and permission details, then confirm or cancel.

    Tip: Before confirming, pay particular attention to the repository name and the scope of the secret to be created.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • SettingClaude Code

    Uncommitted `/config` changes are now clearly reported

    If a settings change made with `/config` or a similar command cannot be written to a file, Claude Code no longer continues silently; it displays the save error below the prompt.

    Usage details

    Who can use it: Claude Code 2.1.327; sessions with a settings file that is corrupted or lacks write permission.

    How to use it: After changing a setting, read the warning below the prompt, fix the file permissions or managed settings source, and apply the change again.

    Tip: If you see a warning, check whether the settings file actually changed before restarting the session.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • APIClaude Code

    MCP server JSON can now be added securely via stdin

    The `claude mcp add-json <name> -` command reads the MCP server’s JSON configuration from stdin. This makes it possible to add a server without exposing access tokens on the command line or in shell history.

    Usage details

    Who can use it: Claude Code 2.1.327 CLI; cannot be used from within a Claude Code session.

    How to use it: Pipe in the JSON by running `cat server.json | claude mcp add-json my-server -`, or connect a secure pipeline to stdin.

    Tip: Pipe JSON from a secret store and delete any temporary file to avoid writing secret values to shell history.

  • CapabilityClaude Code

    `/model` no longer shows the wrong models on custom LLM gateways

    Fixed an issue where `/model` showed the wrong models when `ANTHROPIC_BASE_URL` in project settings pointed to an LLM gateway.

    Usage details

    Who can use it: Claude Code terminal sessions using `ANTHROPIC_BASE_URL`.

    How to use it: Set the gateway address in project settings, then open `/model` to check the model list provided by the server.

    Tip: After changing the gateway, reopen the `/model` screen and verify your selection to avoid using an outdated model list.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • AgentClaude Code

    Plugin agents are preserved after loading and resuming

    Fixed an issue where agents passed with `--agents` disappeared after a plugin was loaded or reloaded, and where a plugin agent was dropped after using the `--resume` picker.

    Usage details

    Who can use it: Claude Code 2.1.327; terminal sessions using plugins and custom agents.

    How to use it: Start custom agents with `--agents` or load them through a plugin; after refreshing the plugin, reopen the session from the `/resume` picker if needed.

    Tip: Keep critical agent definitions under version control outside the plugin as well, so you can recover quickly from loading issues.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • APIClaude Code

    Writes are blocked when project MCP configuration is malformed

    `claude mcp add --scope project` and `claude mcp remove --scope project` now stop rather than overwrite other top-level fields in `.mcp.json` when servers are at the top level and the `mcpServers` key is missing.

    Usage details

    Who can use it: Claude Code 2.1.327; repositories using project-scoped MCP configuration.

    How to use it: If the command stops, move the servers in `.mcp.json` under `mcpServers` in the correct format, then rerun `claude mcp add --scope project ...`.

    Tip: Put the MCP file under version control before modifying it; preserve additional top-level fields such as `$schema`.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    Long replies for unrecognized models now continue up to 128,000 tokens

    Long replies that were cut off at 32,000 output tokens for unrecognized model IDs now continue with a 128,000-token limit.

    Usage details

    Who can use it: Claude Code sessions using a custom gateway or an unrecognized `claude-*` model ID.

    How to use it: Send a request as usual with your custom model ID; if the reply exceeds 32,000 tokens, Claude Code will manage the continuation request automatically.

    Tip: Also check the gateway’s actual output limit; even if the client raises its limit, the provider may enforce a lower one.

  • Plugin / IntegrationClaude Code

    Remote MCP servers are retried in the background after long outages

    After long outages, such as a VPN disconnection, remote MCP servers accessed over HTTPS by hostname will attempt to reconnect in the background for about a day.

    Usage details

    Who can use it: Remote HTTPS MCP servers; especially cloud, self-hosted, and sessions affected by network outages.

    How to use it: Configure the MCP server as usual; if the connection drops, check `/mcp` status and wait for the background reconnection to complete.

    Tip: For short VPN outages, wait for automatic reconnection before deleting and re-adding the server.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude in Chrome

    Timeout for long `browser_batch` calls increased to 150 seconds

    Long `browser_batch` calls in Claude in Chrome now wait up to 150 seconds for a response instead of 90, reducing the chance of completed results being lost to a timeout.

    Usage details

    Who can use it: Claude Code sessions using Claude in Chrome tools.

    How to use it: Run multiple browser actions in `browser_batch`; for tasks that take longer than 150 seconds, split the work into smaller batches.

    Tip: On heavy pages, using two smaller batches instead of one large batch also makes debugging easier.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • AgentClaude Code

    Long `ListAgents` results are paginated

    For accounts with many cloud or Remote Control sessions, long `ListAgents` results are now paginated instead of hiding most sessions.

    Usage details

    Who can use it: Claude Code users managing many sessions or agents.

    How to use it: Open the agent view or flows that use `ListAgents`; if results are paginated, move between pages.

    Tip: Using a naming convention for many background sessions makes it easier to quickly find the right one.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • FeatureClaude Code

    VS Code Explorer gets a command to add selections directly to chat

    Right-clicking files or folders in VS Code Explorer now shows the `Claude Code: Add to Chat` command. Selected paths are added to the chat as `@`-mentions.

    Usage details

    Who can use it: Claude Code VS Code extension; workspaces using VS Code Explorer.

    How to use it: Select files or folders in Explorer, right-click, and run `Claude Code: Add to Chat`.

    Tip: Add only the folders needed for the task to Claude to reduce context and file search time.

  • SettingClaude Code

    `/config` rows identify values overridden by higher-priority settings

    `/config` now displays a note when a setting row is overridden by another settings source. This makes it easier to see why a user setting is having no effect.

    Usage details

    Who can use it: Claude Code 2.1.327; sessions where user, project, or managed settings are applied together.

    How to use it: Open `/config`, read the override note on the row that appears ineffective, and change the higher-priority settings source if needed.

    Tip: If organization policy or managed settings are in place, confirm which source takes precedence before changing a local setting.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • FeatureClaude Code

    `/doctor` starts showing findings before the full report is ready

    Instead of remaining silent until the entire check is complete, `/doctor` now shares findings as it goes.

    Usage details

    Who can use it: Claude Code terminal users.

    How to use it: Run `/doctor` in the terminal and review report lines as they appear; you can fix visible issues without waiting for it to finish.

    Tip: Address errors and configuration warnings first, then review performance information at the end of the report.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • SettingClaude Code

    `/statusline` writes to the settings directory selected by `CLAUDE_CONFIG_DIR`

    Fixed an issue where `/statusline` mistakenly wrote to `~/.claude/settings.json` when `CLAUDE_CONFIG_DIR` was set to a different directory.

    Usage details

    Who can use it: Claude Code installations using a custom `CLAUDE_CONFIG_DIR`.

    How to use it: Set `CLAUDE_CONFIG_DIR` and run `/statusline`; check that the change is written to the settings file in the selected configuration directory.

    Tip: If you use multiple profiles, verify `CLAUDE_CONFIG_DIR` and statusline settings together for each profile.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • SettingClaude Code

    Errors in malformed settings or `.mcp.json` files show the line and column

    For settings or `.mcp.json` files containing invalid JSON, the error message now reports the line and column of the error.

    Usage details

    Who can use it: Claude Code settings and MCP configuration files.

    How to use it: If you see a JSON error at startup or while running a command, go to the specified line and column and fix the comma, quotation mark, or bracket error.

    Tip: Running a JSON linter before changing the file helps prevent Claude Code restart loops.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    Only the first 100 files are searched in prompts with many `@` mentions

    To prevent prompts containing thousands of `@` mentions from hanging before a response, file search is limited to the first 100 mentions.

    Usage details

    Who can use it: Claude Code terminal and related prompt interfaces.

    How to use it: If you need to add more than 100 files, split the prompt across several messages or add files using narrower folder selections.

    Tip: For bulk context, selecting a folder path is more efficient than adding hundreds of files one by one with `@`.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    Resumed sessions are faster with many connector permissions

    A slowdown in resumed sessions for organizations with tens of thousands of connector tool permissions has been fixed.

    Usage details

    Who can use it: Cloud or enterprise Claude Code sessions with many connector permissions configured.

    How to use it: Resume the current session as usual; no additional settings are required. If needed, reopen the same session from the `/resume` list.

    Tip: Regularly removing unused connector permissions helps improve overall session speed, not just the benefits of this fix.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • APIClaude Code

    Managed settings checksum added to gateway hint headers

    When `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1` is enabled, Claude Code adds `x-claude-code-managed-settings-checksum` to gateway hint headers. This header helps the gateway identify the applied managed settings version.

    Usage details

    Who can use it: Claude Code gateway deployments using managed settings; disabled by default.

    How to use it: To enable gateway hint headers, set `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1` in the runtime environment; read the checksum header on the gateway side.

    Tip: If you use gateway caching, use the checksum as a cache invalidation key when settings change.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    The Read tool reports the actual file size for large files

    When a file is 10 MB or larger and too large to read at once, the Read tool now reports its actual size instead of a fixed 512 KB.

    Usage details

    Who can use it: Claude Code Read tool.

    How to use it: Use an appropriate offset and limit in the Read call to split up a large file; refer to the actual size in the warning.

    Tip: For large logs, reduce context and token usage by searching first or reading a narrow range of lines.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • SettingClaude Code

    `/tui classic` added as the name for the non-full-screen interface

    The non-full-screen renderer can now be selected with `/tui classic`. The old `/tui default` name continues to work.

    Usage details

    Who can use it: Claude Code 2.1.327 terminal interface.

    How to use it: Type `/tui classic` in the terminal; `/tui default` is also valid if you want to use the existing default name.

    Tip: Try the `classic` renderer if full-screen behavior causes problems in small or SSH terminals.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü
  • CapabilityClaude Code

    WebFetch now rejects URLs longer than 2,000 characters in advance

    Very long WebFetch URLs are now rejected before permission is requested; the limit is 2,000 characters. This prevents long URLs from blocking compaction or session flow.

    Usage details

    Who can use it: Claude Code WebFetch tool.

    How to use it: For links over 2,000 characters, shorten the query parameters, split the URL into a page or API request, and try again.

    Tip: Removing unnecessary tracking parameters (such as `utm_*`) brings most long URLs below the limit.

    ✓ Verified by official sourceSource: Claude Code 2.1.327 GitHub sürümü

OpenAI · ChatGPT

#2 GPT-6.2 Astra#5 GPT-6.2 Sol#8 GPT-6.1 Terra
Full history (339) →

New in this edition · 28

  • AvailabilityDots in ChatGPT Enterprise, Edu ve Healthcare

    Dots became enabled by default in eligible workspaces on November 18

    The Enterprise admin preview for Dots has ended, and access is now enabled by default in eligible Enterprise, Edu, and Healthcare workspaces. Admins can still opt out, manage access by role, and dots are unavailable in workspaces with UAE data or inference residency.

    Usage details

    Who can use it: Eligible Enterprise, Edu, and Healthcare workspaces; members can use dots unless an admin opts out. Access and computer/app capabilities depend on roles and Agent Security policies.

    How to use it: Review and save the Use dots permission and custom roles at Workspace settings > Permissions & roles > Permissions > Workspace capabilities; users create their own dot from the dots profile.

    Tip: Before enabling it by default, restrict Use dots, local computer, cloud browser, and app action permissions by role.

  • Pricing / PlanDots in ChatGPT Enterprise

    Enterprise dots tiers and free period announced

    A separate pricing page lists monthly prices of 50, 200, and 500 dollars for Standard, Advanced, and Expert dots, respectively, and says they are free through January 13, 2027; billing starts on January 14. The general launch guide, however, says billing starts on November 18, so OpenAI’s two help pages conflict.

    Usage details

    Who can use it: ChatGPT Enterprise workspaces; all enabled dots start at the Standard tier, and admins can move them to a higher tier. Separate Work or Codex jobs started by a dot also count against existing quotas or limits.

    How to use it: Check the dot tier and active status in Admin Console > Dots; confirm the applicable start date with OpenAI’s account team before finalizing a contract or billing plan.

    Tip: Budget dot costs separately from Work/Codex job costs; confirm the free period and billing date in writing.

  • ModelChatGPT Work, Codex ve OpenAI API

    GPT-6.2 Astra and Sol launched in ChatGPT Work, Codex, and API

    OpenAI released GPT-6.2 Astra for difficult professional work and GPT-6.2 Sol for everyday work and agent scale. Astra API prices are lower than those of GPT-6 Astra; Sol retains GPT-6.1 Sol’s prices, and both models support tools, computer use, and multi-agent workflows.

    Usage details

    Who can use it: Plus, Pro, Enterprise, and Edu users can use them in ChatGPT Work and Codex; Pro 500 has Ultrafast for both models, while Pro 200 initially has Ultrafast for Sol only. Developers can use the gpt-6.2-astra and gpt-6.2-sol identifiers in the API; Enterprise and Edu access depends on admin enablement.

    How to use it: Select a model from the model picker in ChatGPT Work or Codex; in the API, set the model field in a Responses request to gpt-6.2-astra or gpt-6.2-sol.

    Tip: Keep the prompt prefix fixed and append new instructions to the end to preserve cache reuse and low latency.

  • CapabilityCodex Cloud

    Codex Cloud preserves conversations and recovery turns during connection drops

    Codex 0.170.0 improved reconnection for idle Cloud conversations after a temporary stream loss and preserves active recovery turns when the client reconnects. This reduces the risk of losing results during long-running cloud tasks.

    Usage details

    Who can use it: Clients using Codex Cloud, version 0.170.0 and later.

    How to use it: Update Codex; if the connection drops, reopen the same Cloud conversation and monitor the status of the active recovery turn.

    Tip: For long-running tasks, save intermediate states and thread IDs to make it easier to resume control after reconnecting.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • Plugin / IntegrationCodex

    Codex fixes MCP OAuth credential race conditions

    Codex 0.170.0 fixes race conditions that could cause MCP OAuth credentials to be overwritten with stale data during cancellation, sign-in, and sign-out. Switching sessions and canceling operations in MCP connections should now be more reliable.

    Usage details

    Who can use it: Codex 0.170.0 and later; MCP servers using OAuth.

    How to use it: Update Codex; when switching MCP accounts, cancel the active call and complete the sign-in flow as usual.

    Tip: Log credential refresh errors securely, but never log access token values.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • CapabilityCodex TUI

    Codex TUI adds WebRTC negotiation for real-time voice calls

    Codex 0.170.0 adds WebRTC negotiation for real-time voice calls in the TUI. This provides a new transport option for setting up voice sessions.

    Usage details

    Who can use it: Codex 0.170.0 and later; TUI real-time voice calls.

    How to use it: Upgrade Codex to 0.170.0 and start a real-time voice call from the TUI; connection negotiation takes place over WebRTC.

    Tip: If you experience voice connection issues on a corporate network, check WebRTC permissions and proxy/firewall rules in advance.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • SettingCodex

    Codex preserves Windows sandbox and child thread permissions more consistently

    Codex 0.170.0 fixes Windows file modification, sandbox permission preservation, and runtime record recovery. It also fixes cases where permission updates and child threads could cause existing sandbox restrictions to be lost.

    Usage details

    Who can use it: Codex 0.170.0 and later; especially users of Windows sandbox and child threads.

    How to use it: Upgrade Codex and review existing sandbox permission profiles; validate tasks that start child threads in a restricted test workspace.

    Tip: After changing security policies, test the permission profile and file access in a small test repository.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • CapabilityCodex

    Codex stops unnecessarily retrying oversized WebSocket requests

    Codex 0.170.0 routes WebSocket requests that exceed the size limit to HTTP without unnecessary retries. This transport resilience fix reduces stalls and delays for large requests.

    Usage details

    Who can use it: Codex 0.170.0 and later; large Responses/WebSocket requests.

    How to use it: Update Codex; if you use a custom WebSocket message size setting, respect the limits advertised by the server.

    Tip: Monitor HTTP fallback requests at the gateway as well; don’t rely only on WebSocket metrics.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • Plugin / IntegrationCodex

    Codex experimentally supports MCP servers with the same name in different environments

    Codex 0.170.0 adds experimental support for keeping MCP servers with the same name in different working environments separate. This lets local, remote, or cloud environments use the same logical server name in their configurations.

    Usage details

    Who can use it: Codex 0.170.0 and later; experimental MCP environment configurations.

    How to use it: Update Codex and use the same server name in environment-specific MCP configurations; define each environment’s endpoint and credentials separately.

    Tip: Manage MCP configuration with version control and separate secrets to prevent incorrect credentials or data access across environments.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • CapabilityCodex

    Codex removes duplicate file search results and fixes snippets for queries with spaces

    Codex 0.170.0 removes duplicate file search results from overlapping roots. It also fixes an issue that caused thread search snippets to appear incorrectly for queries containing spaces.

    Usage details

    Who can use it: Codex TUI and clients using thread/file search, version 0.170.0 and later.

    How to use it: Update Codex and use file or thread search in the TUI with regular queries; no additional settings are needed.

    Tip: If you scan the same file tree from multiple roots, compare the result count and search context with the new version before upgrading.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • Plugin / IntegrationCodex

    Codex refreshes authentication headers for OpenTelemetry HTTP exporters

    Codex 0.170.0 adds automatically refreshed authentication headers for OpenTelemetry HTTP exporters. This reduces the need for manual restarts in observability setups that use expiring credentials.

    Usage details

    Who can use it: Codex 0.170.0 and later; setups using the OpenTelemetry HTTP exporter.

    How to use it: Upgrade Codex to 0.170.0 and configure your exporter credential provider to support refreshable headers.

    Tip: Monitor header refresh and authorization denial events as separate alerts in your monitoring backend.

    ✓ Verified by official sourceSource: OpenAI Codex sürüm notları
  • APIOpenAI Agents SDK for Python

    Agents SDK checkpoint schema 1.22 and Pydantic context serialization added

    Agents SDK 0.25.0 writes RunState checkpoint schema 1.22 to support resuming nested agent-tool approvals, and preserves Pydantic context defaults and in-place changes during JSON serialization. Older 0.24.0 workers cannot read checkpoints written by the new writer.

    Usage details

    Who can use it: Agents SDK applications that share or resume checkpoints.

    How to use it: Upgrade checkpoint readers first; provide context_serializer, context_deserializer, or context_override for process-local fields, and implement version routing.

    Tip: Use a reader-before-writer deployment order and test the rollback plan from 1.22 to 1.21.

  • RetiredOpenAI Agents SDK for Python

    Agents SDK for Python removed the Runloop sandbox provider

    Agents SDK for Python 0.25.0 removed the beta Runloop sandbox provider, its related extra, and exports. Applications using Runloop need to switch to another supported provider or stay on 0.24.0 until migration is complete.

    Usage details

    Who can use it: Sandbox applications upgrading to 0.25.0; the Runloop integration is no longer available.

    How to use it: Preserve any required workspace data; if switching to Modal, install openai-agents[modal] and update the SandboxRunConfig client and options. Old Runloop session IDs and snapshots cannot be resumed directly in Modal.

    Tip: Do not start writing checkpoints with the new provider until reader workers and data migration are complete.

  • ModelAstra for Law

    Astra for Law updated; all Enterprise customers can request access

    OpenAI released an updated version of the GPT-6.2 Astra-based Astra for Law configuration and expanded its legal search index. Any ChatGPT Enterprise customer can request ChatGPT and API access through the Trusted Access Program.

    Usage details

    Who can use it: ChatGPT Enterprise customers; access is subject to Trusted Access approval, and API use requires a separate request.

    How to use it: Request Astra for Law access through the Trusted Access Program in the admin console; after approval, select the gpt-6.2-astra-law model in ChatGPT or use OpenAI’s API access form.

    Tip: Have an expert review source documents and citations before using legal output directly.

    ✓ Verified by official sourceModels: GPT-6.2 Astra, Astra for Law, gpt-6.2-astra-lawSource: OpenAI: Astra for Law hukuki zekâda öne çıkıyor
  • Plugin / IntegrationChatGPT for Sales

    ChatGPT for Sales and more than 30 partner plugins introduced

    ChatGPT for Sales introduces a new Sales plugin that handles account planning, meeting preparation, follow-up messages, and pipeline risks in a single workflow. OpenAI also offered more than 30 new or improved partner plugins alongside sales context and customer tools.

    Usage details

    Who can use it: Sales teams using ChatGPT Work; plugin availability varies by plan, region, and workspace settings. Business and Enterprise admins can enable the plugin for their team or set it as a default via Admin Console > Plugins.

    How to use it: Install the Sales plugin; if you’re a Business or Enterprise admin, open the workspace and Plugins section in Admin Console and enable Sales and the necessary partner plugins, then connect account or meeting context.

    Tip: First connect CRM, meeting, and email context and ask for account prioritization; review writing or sending actions before approving them.

  • CapabilityOpenAI Agents SDK for Python

    Agents SDK added optional recovery for handoff errors

    Agents SDK for Python 0.25.0 can route handoff failures to an optional error recovery function. If no option is provided, the previous exception propagation behavior is preserved.

    Usage details

    Who can use it: openai-agents 0.25.0 and later; Runner handoff flows.

    How to use it: Provide failure_error_function in the Runner handoff definition to route handoff errors to your application’s recovery path; leave the field empty if you want the default exception behavior.

    Tip: Use an idempotent fallback and a clear user-facing error state in the recovery function instead of unlimited retries.

  • SettingOpenAI Agents SDK for Python

    Agents SDK changed model defaults and Chat Completions behavior

    In 0.25.0, the default for OpenAIResponsesCompactionSession is now gpt-6-luna, and model settings better match the selected provider model. For GPT-6 Sol and GPT-6 Luna, Chat Completions tool requests default to non-reasoning when effort is not specified.

    Usage details

    Who can use it: Agents SDK for Python 0.25.0 and later.

    How to use it: To keep the old compaction model, provide OpenAIResponsesCompactionSession(model="gpt-4.1"); to use the new default, use it without specifying a model.

    Tip: Before upgrading in production, compare token cost, latency, and reasoning behavior using a fixed evaluation set.

    ✓ Verified by official sourceModels: gpt-6-luna, gpt-6-solSource: OpenAI Agents SDK for Python 0.25.0 sürümü
  • Plugin / IntegrationOpenAI Agents SDK for Python

    Agents SDK made MCP v2 Streamable HTTP requests concurrent

    Agents SDK for Python 0.25.0 supports concurrent requests over MCP v2 Streamable HTTP; MCP v1 requests remain serialized. This can reduce waiting in agents that make parallel tool calls.

    Usage details

    Who can use it: Python Agents SDK applications using MCP v2 Streamable HTTP.

    How to use it: Upgrade to openai-agents 0.25.0 and compatible MCP dependencies, and use an MCP v2 Streamable HTTP server; enable call retries only for safe and idempotent operations.

    Tip: Separately test server-side rate limits and tool side-effect safety for concurrent calls.

  • APIOpenAI Agents SDK for Python

    Agents SDK opens helper utilities for third-party session adapters

    Agents SDK for Python 0.25.0 exports helpers for writing third-party session adapters. This lets applications connect their own session stores more directly to the SDK’s session flow.

    Usage details

    Who can use it: Python developers using openai-agents 0.25.0 or later.

    How to use it: Upgrade to 0.25.0 or later with pip install -U openai-agents, then connect your own session backend using the SDK’s third-party adapter APIs.

    Tip: Test versioning, concurrent writes, and restart behavior in the adapter layer from the outset.

  • APIOpenAI Go SDK

    Go SDK adds 2/8-hour background windows and GPT-6.2 definitions

    openai-go 3.101.0 adds 2-hour and 8-hour background completion windows for Responses, along with the GPT-6.2 Astra, Sol, and Astra Law model IDs.

    Usage details

    Who can use it: Developers using the Responses API with Go; background windows are tied to models covered by the public beta.

    How to use it: Upgrade the Go module to v3.101.0 or later; send background_options.completion_window as 2h or 8h in a Responses request and select one of the new model constants.

    Tip: Preserve discounts and queue priority in follow-up requests for the same workflow by keeping the model, project, service tier, and completion window values unchanged.

    ✓ Verified by official sourceModels: GPT-6.2 Astra, GPT-6.2 Sol, gpt-6.2-astra-lawSource: OpenAI Go SDK 3.101.0 sürümüSource: Responses arka plan modu
  • APIOpenAI Go SDK

    Go SDK adds types for Managed Agents budget events

    openai-go 3.102.0 enables typed handling of Managed Agents budget events. This lets Go applications process stream budgets more safely and with compile-time checks.

    Usage details

    Who can use it: Developers using Managed Agents with Go.

    How to use it: Upgrade the Go module to v3.102.0 or later and parse Managed Agents stream events using the new types.

    Tip: Manage interrupted tasks by adding automatic pauses or user approval when approaching budget exhaustion.

    ✓ Verified by official sourceSource: OpenAI Go SDK sürümleri
  • APIOpenAI Node SDK

    Node SDK adds background windows, GPT-6.2 definitions, and scheduling reasons

    openai-node 7.58.0 adds 2h/8h background windows for Responses, GPT-6.2 model IDs, and the schedule pause reasons field. Scheduling pause reasons are now supported in Node, following their introduction in the Python, Go, and Ruby SDKs on November 16.

    Usage details

    Who can use it: Developers using the Responses and Agents APIs with Node/TypeScript; background windows are subject to the public beta.

    How to use it: Upgrade the openai package to 7.58.0 or later; use 2h/8h for background_options.completion_window in a Responses request and read the pause-reason field in scheduling responses.

    Tip: Reflect the pause reason value directly in the status screen to show users why scheduled tasks stopped.

    ✓ Verified by official sourceModels: GPT-6.2 Astra, GPT-6.2 Sol, gpt-6.2-astra-lawSource: OpenAI Node SDK 7.58.0 sürümüSource: Responses arka plan modu
  • APIOpenAI Node SDK

    Node SDK adds types for Managed Agents budget events

    openai-node 7.59.0 adds typed support for Managed Agents budget events. TypeScript applications can handle budget events with compile-time safety.

    Usage details

    Who can use it: Developers using Managed Agents with Node or TypeScript.

    How to use it: Upgrade the openai package to 7.59.0 or later and use the budget types in Managed Agents stream events.

    Tip: Send budget events to your observability system and request user approval before starting another turn as the limit approaches.

    ✓ Verified by official sourceSource: OpenAI Node SDK sürümleri
  • APIOpenAI Python SDK

    Python SDK adds 2/8-hour background windows and GPT-6.2 definitions

    openai-python 3.53.0 adds 2-hour and 8-hour completion window options for Responses background requests. The same release recognizes the GPT-6.2 Astra, GPT-6.2 Sol, and GPT-6.2 Astra Law model IDs.

    Usage details

    Who can use it: Developers using the Responses API with Python; 2h and 8h windows depend on the public beta and eligible models.

    How to use it: Upgrade the openai package to 3.53.0 or later; set background=true and background_options.completion_window to 2h or 8h in a Responses request, and use one of the new model IDs.

    Tip: Choose the 8h window for overnight document reviews and the 2h window for long analyses to be completed the same day.

    ✓ Verified by official sourceModels: GPT-6.2 Astra, GPT-6.2 Sol, gpt-6.2-astra-lawSource: OpenAI Python SDK 3.53.0 sürümüSource: Responses arka plan modu
  • APIOpenAI Python SDK

    Python SDK adds typed budget events for Managed Agents

    openai-python 3.54.0 added typed SDK support for budget events in Managed Agents streams. This makes it easier for applications to handle budget status using client types instead of parsing raw events.

    Usage details

    Who can use it: Developers using Managed Agents with Python.

    How to use it: Upgrade the openai package to 3.54.0 or later and handle budget events in the Managed Agents stream through SDK-generated typed events.

    Tip: Connect budget events to logging and user capacity alerts; don’t wait for the final error.

    ✓ Verified by official sourceSource: OpenAI Python SDK sürümleri
  • APIOpenAI Ruby SDK

    Ruby SDK adds 2/8-hour background windows and GPT-6.2 definitions

    openai-ruby 0.126.0 added 2h and 8h completion window options for Responses background requests, along with the GPT-6.2 Astra, Sol, and Astra Law model IDs.

    Usage details

    Who can use it: Developers using the Responses API with Ruby; the 2h and 8h options are in public beta and depend on eligible models.

    How to use it: Upgrade the gem to 0.126.0 or later; set background_options.completion_window to 2h or 8h in the Responses creation call and provide the new model ID.

    Tip: Move long-running tasks to the 8h window and track the result with a webhook or polling; don’t keep the connection open while the user waits.

    ✓ Verified by official sourceModels: GPT-6.2 Astra, GPT-6.2 Sol, gpt-6.2-astra-lawSource: OpenAI Ruby SDK 0.126.0 sürümüSource: Responses arka plan modu
  • APIOpenAI Ruby SDK

    Ruby SDK adds typed budget events for Managed Agents

    openai-ruby 0.127.0 made Managed Agents budget events typed. Ruby applications can now handle these events with less need for generic hash parsing.

    Usage details

    Who can use it: Developers using Managed Agents with Ruby.

    How to use it: Upgrade the gem to 0.127.0 or later and handle budget events in the Managed Agents stream using the new typed event objects.

    Tip: Connect budget events to the task progress screen and show users their remaining capacity.

    ✓ Verified by official sourceSource: OpenAI Ruby SDK sürümleri
  • AvailabilityChatGPT for Academic Researchers

    Academic researcher program expands to up to 200,000 people

    OpenAI is expanding the ChatGPT for Academic Researchers program to up to 200,000 researchers and will award 50,000 new grants to eligible applicants in the new round. Each grant provides the recipient and up to two colleagues with 12 months of free ChatGPT Plus access.

    Usage details

    Who can use it: Academic researchers in eligible regions outside China and Hong Kong; new applications open on November 23, 2026, and are processed on a first-come, first-served basis for the first 50,000 eligible applications.

    How to use it: Apply at chatgpt.com/academic-researchers/apply on November 23; those already on the waitlist will receive activation instructions by email after eligibility verification.

    Tip: Prepare proof of academic affiliation and eligible publications before applying; check that program data is not used for model training by default.

Google · Gemini

#3 Gemini 4 Carbon#7 Gemini 4 Hydrogen#9 Gemini 4 Argon
Full history (244) →

New in this edition · 9

  • APIGoogle Cloud CLI / Agent Registry

    Agent Registry CLI encryption and revision download commands reach GA

    The gcloud agent-registry skill-encryption-spec and gcloud agent-registry skills revisions download commands in Google Cloud CLI have been upgraded to GA. The same release also fixed uploading a local ZIP archive as the first skill revision when using skills create --payload.

    Usage details

    Who can use it: Available in Google Cloud CLI 594.0.0 and later; usable in Google Cloud projects that manage Agent Registry skills.

    How to use it: Update the CLI to 594.0.0 or later; use gcloud agent-registry skill-encryption-spec to specify encryption and gcloud agent-registry skills revisions download to download revisions.

    Tip: In CI/CD deployments, download and validate skill revisions before proceeding; this catches incorrect or incomplete ZIP uploads early.

  • AvailabilityAlphaGenome

    Model Garden allowlist requirement removed for AlphaGenome

    AlphaGenome no longer requires a project allowlist. Any Google Cloud project can view and deploy the alphagenome-003 or alphagenome-004 model through Model Garden without submitting an access request.

    Usage details

    Who can use it: Available in Model Garden for all projects on Google Cloud; deployment requires a compatible A100 or H100 GPU, quota, and paid infrastructure.

    How to use it: In Google Cloud Console, open Model Garden > AlphaGenome, select a compatible version, and configure the deployment options; you no longer need to submit an allowlist access form.

    Tip: Verify GPU quota in the target region before deployment to meet the 1 MB context and high GPU memory requirements.

    ✓ Verified by official sourceModels: alphagenome-003, alphagenome-004Source: Gemini Enterprise Agent Platform sürüm notlarıSource: AlphaGenome dokümantasyonu
  • CapabilityBigQuery generative AI functions

    Optimized mode for BigQuery AI.IF and AI.CLASSIFY reaches GA

    Optimized mode is now generally available for BigQuery’s managed AI.IF and AI.CLASSIFY functions. By creating a lightweight distilled model from sample data, BigQuery can reduce Gemini calls, token consumption, and query latency on large tables.

    Usage details

    Who can use it: Available in projects using BigQuery managed AI functions, especially with datasets containing thousands or billions of rows.

    How to use it: Provide embeddings in an AI.IF or AI.CLASSIFY call and use optimization_mode => MINIMIZE_COST; monitor the result in BigQuery Console under Jobs explorer > Job information > Gen AI function optimizations.

    Tip: Provide approximately 3,000 or more rows and consistently sized embeddings from the same model; optimization may fail with small or imbalanced samples.

  • APIBigQuery AI.RANK

    BigQuery AI.RANK natural-language relevance ranking reaches GA

    AI.RANK is now generally available for ranking text in a specified column of a table by how well it answers a natural-language query. The function assigns each row a relevance score and rank position, and preserves all input columns in the query result.

    Usage details

    Who can use it: Available in BigQuery projects with the Discovery Engine Viewer role; it does not index results, and costs increase with the number of rows processed.

    How to use it: In SQL, call AI.RANK(TABLE table, text_column, search query, top_n => N); enable the Discovery Engine API and grant the required roles/discoveryengine.viewer role.

    Tip: To limit costs, first narrow down candidates with AI.SEARCH or VECTOR_SEARCH, then rerank only those candidates with AI.RANK.

  • APIBigQuery AI.SEARCH

    BigQuery rerank for AI.SEARCH result ranking reaches GA

    A rerank option that reorders results based on how well they answer the search query has been added to BigQuery’s AI.SEARCH function as GA. It can be used to improve the relevance ranking of results returned by semantic or hybrid search.

    Usage details

    Who can use it: Available for tables in BigQuery that use AI.SEARCH and have autonomous embedding generation configured.

    How to use it: Enable the new rerank option when running your AI.SEARCH query; narrow candidates with semantic search and use the reranking results in the BigQuery query output.

    Tip: Keep top_k at a reasonable value in the initial search and rerank only the candidate set; this provides a better balance between accuracy and cost.

  • APIGoogle Cloud CLI / Model Garden

    --system-labels no longer ignored in Model Garden deployments

    Google Cloud CLI 594.0.0 fixed an issue where the --system-labels flag in the gcloud ai model-garden models deploy command was ignored for Model Garden and Hugging Face models. System labels specified during deployment are now applied.

    Usage details

    Who can use it: Available in Google Cloud CLI 594.0.0 and later; usable in projects deploying Model Garden or Hugging Face models.

    How to use it: After updating the CLI, run gcloud ai model-garden models deploy --model MODEL --system-labels=KEY=VALUE and verify that the labels are present on the deployment resource.

    Tip: For tracking and cost reporting, apply team, environment, and application labels at deployment time using a standardized vocabulary.

  • Plugin / IntegrationPersonalized Service Health

    Personalized Service Health adds support for Google Antigravity

    Enterprise Personalized Service Health has added Google Antigravity to its list of supported products. This allows incident and status notifications for Antigravity to be included in service health monitoring workflows via the product ID.

    Usage details

    Who can use it: Personalized Service Health is available to organizations using Google Cloud; Google Antigravity can be identified by product ID hZpxrZf6wIdajCN6y30V.

    How to use it: When creating a new incident alert or policy in Personalized Service Health, select Google Antigravity as the product, or use product ID hZpxrZf6wIdajCN6y30V in API and log matching.

    Tip: Use the stable product ID rather than the display name in programmatic alerts; your integration won’t be affected by renames.

  • AgentGemini Enterprise Agent Platform

    Deferred tier reaches GA for ADK autonomous agent scheduling

    Deferred tier, which shifts latency-tolerant autonomous agent workloads to off-peak hours, is now generally available. Support now includes ADK agents built with google-adk v2.10.0 and later, in addition to Deep Research Agent, and provides a 50 percent discount on model inference charges.

    Usage details

    Who can use it: Available on Gemini Enterprise Agent Platform across Agent Runtime, Cloud Run, and GKE for ADK v2.10.0 and later agents and Gemini Deep Research Agent.

    How to use it: For a model that uses the Interactions API, set ServiceTier.DEFERRED in the ADK RunConfig; for Deep Research Agent, create tasks with service_tier=deferred, background=true, and store=true.

    Tip: Move non-real-time batch research, security scanning, and code transformations to Deferred tier to reduce both 429 risk and costs.

    ✓ Verified by official sourceModels: gemini-4-carbon, gemini-4-argon, gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-4-carbon-cyber-permissive, gemini-4-argon-cyber-permissiveSource: Gemini Enterprise Agent Platform sürüm notlarıSource: Otonom ajan zamanlaması
  • APIInteractions API

    Deferred tier for asynchronous Gemini calls in the Interactions API reaches GA

    service_tier=deferred is now generally available on the Interactions API's global endpoint. Calls can be queued with a 24-hour completion target, status is retained for 7 days, and a 50 percent discount applies compared with Standard PayGo.

    Usage details

    Who can use it: Available for first- and third-party agents on the global endpoint; usable with supported Gemini 3 and Gemini 4 models through REST or the Google Gen AI SDK.

    How to use it: Send service_tier=deferred, background=true, and store=true in a POST /v1/projects/PROJECT_ID/locations/global/interactions request; poll, stream, or cancel the result using the interaction ID.

    Tip: Chain multiple turns with previous_interaction_id; 7-day status retention lets you continue long-running research and evaluation workflows without restarting them.

    ✓ Verified by official sourceModels: gemini-4-carbon, gemini-4-argon, gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-4-carbon-cyber-permissive, gemini-4-argon-cyber-permissiveSource: Gemini Enterprise Agent Platform sürüm notlarıSource: Deferred tier kullanımı

Meta · Meta AI

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

New in this edition · 6

  • FeatureInstagram

    Instagram Content Protection brings original video protection to eligible creators

    Eligible Instagram creators who meet the requirements and join the program can automatically scan for visual matches to their original videos on Instagram and Facebook, then track matches and choose to hide them, add a credit link to the original creator, or release them. The tool works with new videos and eligible older videos.

    Usage details

    Who can use it: Eligible Instagram creators who meet the originality and integrity criteria; access notifications appear in the Activity feed, and creators can also apply.

    How to use it: On Instagram, go to Tools > Your professional tools > Content protection from the Home tab. If you don’t have access, apply at https://www.instagram.com/apply-content-protect; after joining, new original videos are protected automatically, and eligible older videos can be selected manually.

    Tip: Use the allow list for your second account, business partner, or an authorized brand to prevent legitimate reposts from being flagged as false matches.

  • AvailabilityMuse

    Muse launches in Australia on iOS, Android, web, and Mac

    Muse became available in Australia on November 17. Connections to Gmail, Google Calendar, Instagram, and OpenTable, along with local connector experiences such as Canva, Domain, Kogan, MECCA, Officeworks, realestate.com.au, SWEAT, ShopBack, and Ticketek, are rolling out gradually.

    Usage details

    Who can use it: Australia; iOS, Android, the muse.ai website, and the Mac app. Basic use is free for most users, with subscription plans for higher usage; support for AI glasses is coming soon.

    How to use it: In Australia, install the Muse app or visit muse.ai and sign in. To connect a service, go to Settings > Connectors > Connect next to the relevant service > Continue.

    Tip: During initial setup, connect lower-risk services such as Google Calendar and Gmail first, and keep approval prompts enabled for connectors that make purchases or reservations.

  • CapabilityInstagram

    Instagram Content Protection is testing spoken-word matching in AI-altered videos

    Meta said it will start showing copies in Content Protection that match the original video’s spoken words, even if the visuals have been changed with AI. The initial test will involve a small group of creators; access is planned to expand to creators registered on Instagram and Facebook.

    Usage details

    Who can use it: Eligible Instagram and Facebook creators enrolled in Content Protection; the initial small-group test will begin in the weeks after the announcement, and broader access is not yet guaranteed.

    How to use it: First, join the program in Instagram under Home > Tools > Your professional tools > Content protection. Once your account is included in the test, review transcript or spoken words matches alongside visual matches in the same match list.

    Tip: Use clear speech and clean audio in the original video to catch copies with fully altered visuals; check view and follower data in the match details before deciding.

  • Plugin / IntegrationMuse

    The Canva connector in Muse is now available to users worldwide

    Meta announced that Canva is available worldwide as a Muse connector. Users can now ask Muse to design the visuals or content they need with Canva.

    Usage details

    Who can use it: Muse users worldwide; the Canva experience may not be visible everywhere depending on regional and phased rollout.

    How to use it: In Muse, open Settings > Connectors, tap Connect next to Canva, review the information and authorize with Continue; then give Muse a design task.

    Tip: Clearly specify the output type, dimensions, target channel, and brand constraints in your request; have Muse ask for approval before completing the Canva action.

  • CapabilityInstagram

    Instagram tightens recommendation filters in India with AI

    Meta has strengthened AI-based detection to more proactively exclude sexually explicit and violent content that does not comply with the Recommendations Guidelines from Explore, Reels, and Feed recommendations in India. The company says reports of recommended content have fallen by about half compared with the previous year.

    Usage details

    Who can use it: Instagram users in India; Explore, Reels, Feed, Search, and recommended account surfaces.

    How to use it: No action is needed to enable this change. To personalize recommendations, adjust Sensitive Content Control or Political Content Control under Settings and activity > Content preferences in Instagram; for an individual recommendation, select the three-dot menu > Not interested.

    Tip: Marking an unwanted recommendation as Not interested right away helps Instagram tailor subsequent recommendations to your preferences.

  • APIThreads API

    Threads API updates image size limits with automatic resizing

    The Threads API image specifications have been updated: images larger than 4096 × 4096 pixels are automatically resized to fit; images with an aspect ratio of 2:1 or wider are reduced to approximately 16.8 megapixels.

    Usage details

    Who can use it: Integrations using the Threads API; official image formats are JPEG and PNG, and the 8 MB file limit remains unchanged.

    How to use it: Before uploading an image, resize it on the client side to meet the 4096 × 4096 limit and approximately 16.8 MP limit; then use the existing media container and publishing workflow.

    Tip: Resizing images in advance reduces quality loss and unexpected cropping that could result from Meta’s automatic resampling.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 5

  • AgentKimi Code

    A background notification no longer accidentally stops a Tower task

    Kimi Code 2.6.3 fixed a bug that caused a background notification to inadvertently stop a Tower task. Tower tasks are no longer terminated accidentally when this type of notification arrives.

    Usage details

    Who can use it: Users who use Kimi Code Tower tasks; the release notes do not specify plan or regional restrictions.

    How to use it: Upgrade Kimi Code to 2.6.3, start a Tower task, and check that the task continues running when a background notification arrives.

    Tip: For long Tower tasks, monitor the notification stream and task status together; if you experienced unexpected stops before 2.6.3, verify the version first.

  • CapabilityKimi Code

    Foreground Shell output no longer repeats, and the loading state no longer gets stuck

    Kimi Code 2.6.3 fixed a bug that caused foreground Shell command output to repeat and the loading state to remain open. The command result can now be displayed once and return to a completed state.

    Usage details

    Who can use it: Users who update to Kimi Code 2.6.3; the release notes do not specify operating system or plan restrictions.

    How to use it: Run kimi upgrade in the CLI and verify that kimi --version reports 2.6.3.

    Tip: For long or multiline Shell commands, wait for the results panel to close; if you see repeated output, verify the version first.

  • CapabilityKimi Code

    Session restore and branching no longer start a new turn after a background notification

    Kimi Code 2.6.3 fixed an issue where a background notification could incorrectly start a new response while restoring a session or creating a session branch. Resume and fork flows no longer generate unexpected turns.

    Usage details

    Who can use it: Users who restore sessions or use fork in Kimi Code; the release notes do not specify plan or regional restrictions.

    How to use it: Upgrade Kimi Code to 2.6.3, reopen an existing session with /sessions or create a branch with /fork, and check that no unintended new turn starts after a background notification.

    Tip: In long sessions, check the turn count after resume or fork before sending a new user message.

  • Plugin / IntegrationKimi Code

    The Windows WebBridge daemon can now be updated while running

    Kimi Code 2.6.3 fixed an issue that prevented the WebBridge daemon running on Windows from being updated. This makes the upgrade flow more reliable for installations that use WebBridge.

    Usage details

    Who can use it: Kimi Code users who use WebBridge on Windows; the release notes do not specify plan or regional restrictions.

    How to use it: Upgrade Kimi Code to 2.6.3 with kimi upgrade, then verify the version with kimi --version and rerun a short task that uses WebBridge.

    Tip: After a Windows update, verify the WebBridge connection with a small test task; do not confuse the old daemon version with the new CLI.

  • CapabilityKimi Code

    The external Stop hook now runs at the end of every response turn

    Kimi Code 2.6.3 fixed an issue that caused the external Stop hook—which is expected to be called when each response turn completes—to be skipped. External processes connected to the hook are now triggered more reliably at the end of a turn.

    Usage details

    Who can use it: Kimi Code users who configure an external Stop hook; the release notes do not specify plan, regional, or operating system restrictions.

    How to use it: Upgrade Kimi Code to 2.6.3, complete a response turn with your configured external Stop hook, and observe whether the external command runs.

    Tip: Write the hook’s output to a temporary log file to verify that it is triggered every time across multiple turns.

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 14

  • AvailabilityRunware

    Qwen4-Max is now available on Runware

    The official Qwen account announced that Qwen4-Max has arrived on Runware. The accompanying Runware announcement lists text, image, and video input, a 1 million-token context window, adjustable reasoning effort, function calling, and JSON Schema output.

    Usage details

    Who can use it: Runware users; access and quotas depend on Runware platform terms.

    How to use it: Open Qwen4-Max in the Runware model catalog and send a request through the platform’s API or workspace flow.

    Tip: Use JSON Schema for tasks that require structured output, and adjust reasoning effort to balance cost and depth.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Qwen’in Runware duyurusu
  • AvailabilityQwen4-Max for Enterprise

    Qwen4-Max opens a pilot access program for enterprise teams

    Alibaba Cloud has made Qwen4-Max available for pilot access, allowing enterprise teams to plan its use in their own workflows. This is not a separate model version; access, features, and pricing vary by region.

    Usage details

    Who can use it: Enterprise teams; access and pricing terms vary by region; contact the Alibaba Cloud team to arrange a pilot.

    How to use it: Contact the Alibaba Cloud team to plan a Qwen4-Max pilot for your workflow and request the access terms for your region.

    Tip: Before starting the pilot, define acceptable quality, latency, and per-task cost metrics.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Alibaba Cloud’un Qwen4-Max Enterprise duyurusu
  • AvailabilityLebrelAI

    Qwen4-Max goes live on LebrelAI for agentic coding

    The official Qwen account announced that Qwen4-Max is live on LebrelAI. The LebrelAI post says it offers the model at a price 7 percent below Alibaba’s list price, with signed response receipts.

    Usage details

    Who can use it: LebrelAI users; according to the partner announcement, pricing is 7 percent below list price, but plan and quota details depend on the platform.

    How to use it: Open Lebrel.ai and select Qwen4-Max in the agentic coding workflow; complete the platform’s own access and payment steps.

    Tip: Use the signed receipt feature to compare trial and production results in a traceable way.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Qwen’in LebrelAI duyurusu
  • AvailabilityYou.com Inference Router

    Qwen4-Max added to You.com Inference Router

    The official Alibaba Cloud and Qwen accounts announced that Qwen4-Max is live on You.com Inference Router. The announcement mentions a 1 million-token context window, text, image, and video input, and long-running agent tasks.

    Usage details

    Who can use it: You.com users; access depends on You.com plan and quota terms for Inference Router.

    How to use it: Open `https://you.com/product/inference-router/models/qwen4-max` and select Qwen4-Max to use it through Inference Router.

    Tip: For long-context tasks, provide the model only with the files it needs; compare the router’s model and cost choices for each task.

  • CapabilityQoder

    Qoder 0.5.14 fixes a model validation issue

    Qoder 0.5.14 was released as a minor maintenance update that fixes an error affecting the validation of some models.

    Usage details

    Who can use it: Qoder 0.5.14 users; the fix is included in the version update.

    How to use it: Update Qoder to 0.5.14 or later, then retry model selection or validation.

    Tip: If you use a custom model, revalidate its settings once after updating.

    ✓ Verified by official sourceSource: Qoder sürüm notları
  • Pricing / PlanSophea API

    Qwen4-Max is free and unlimited on Sophea API through November 23

    Alibaba Cloud announced a free, unlimited-use promotion for Qwen4-Max on KieferLab’s Sophea API through November 23, 2026. This is a third-party platform promotion, not a permanent change to Alibaba Model Studio pricing.

    Usage details

    Who can use it: platform.sophea.ai; the announced free, unlimited period runs through 2026-11-23; other terms were not specified.

    How to use it: Open `https://platform.sophea.ai/`, select Qwen4-Max through Sophea API, and complete the platform’s access steps.

    Tip: Don’t assume the promotion is a permanent price; revisit your usage and cost limits before November 23.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Alibaba Cloud’un Sophea API duyurusu
  • CapabilityQoder

    Qoder 0.5.13 recovers from context limit errors through compaction

    Context limit errors in some models are now recognized correctly, and the recovery flow works with the expected conversation compaction.

    Usage details

    Who can use it: Qoder 0.5.13 users; behavior depends on the model’s context limit and conversation length.

    How to use it: Update Qoder to version 0.5.13 or later; in a conversation that reaches the context limit, try continuing after compaction.

    Tip: For very long tasks, save important intermediate outputs to a file before compaction.

    ✓ Verified by official sourceSource: Qoder sürüm notları
  • CapabilityQoder

    Qoder 0.5.13 fixes session startup with invalid subagent model settings

    Qoder fixed an issue where an invalid model setting or configuration in some subagents prevented the main conversation from starting or continuing.

    Usage details

    Who can use it: Qoder 0.5.13 users; the fix comes with the version update.

    How to use it: Update Qoder to version 0.5.13 or later and restart the conversation with the subagent that has the invalid model setting.

    Tip: For custom subagent configurations, compare model IDs against the current Qoder catalog.

    ✓ Verified by official sourceSource: Qoder sürüm notları
  • CapabilityQoder

    Qoder 0.5.13 keeps context usage up to date after model changes

    An issue where old settings continued to be used when the model or context window changed, resulting in incorrect context calculations, has been fixed.

    Usage details

    Who can use it: Qoder 0.5.13 users.

    How to use it: Update Qoder to version 0.5.13 or later; change the model or context window in a conversation, then check the usage indicator again.

    Tip: After changing models, evaluate the remaining context percentage against the new model’s context window.

    ✓ Verified by official sourceSource: Qoder sürüm notları
  • Plugin / IntegrationNovita AI

    Qwen4-Max available on Novita AI from day one

    The official Qwen account announced that Qwen4-Max is live on Novita AI from day one. The Novita announcement highlights the model’s native vision, 1 million token context, stronger tool use, and long task flows.

    Usage details

    Who can use it: Novita AI users; access, plan, and quota terms are determined by Novita AI.

    How to use it: Sign in to Novita AI, select Qwen4-Max from the model catalog, and start a chat or API call through the platform’s own flow.

    Tip: For long agent tasks, monitor how tool calls and intermediate outputs affect the provider quota.

    ✓ Verified by official sourceModels: Qwen4-MaxSource: Qwen’in Novita AI duyurusu
  • Plugin / IntegrationOpenRouter

    Qwen4-Max available through OpenRouter from day one

    The official Qwen account announced that Qwen4-Max is live on OpenRouter. OpenRouter documentation lists the `qwen/qwen4-max` model ID, the Chat Completions endpoint, and options for reasoning, tool calling, and structured output.

    Usage details

    Who can use it: Developers with an OpenRouter account and API key; the model is not available through EU in-region routing.

    How to use it: Use your OpenRouter API key in the `Authorization: Bearer` header and set the `model` field to `qwen/qwen4-max` in a `POST https://openrouter.ai/api/v1/chat/completions` request.

    Tip: If EU-only regional routing is required, check OpenRouter’s EU in-region routing restriction for this model in advance.

  • CapabilityQoder

    Qoder 0.5.13 fixes missing campaign reward notifications

    A fix addresses an issue that prevented some users from receiving notifications to claim campaign benefits.

    Usage details

    Who can use it: Qoder 0.5.13 users; campaign eligibility and benefits may vary by account.

    How to use it: Update Qoder to version 0.5.13 or later and check your account notifications again.

    Tip: If a campaign notification doesn’t appear, also check your account eligibility and notification permissions.

    ✓ Verified by official sourceSource: Qoder sürüm notları
  • CapabilityQoder

    Qoder 0.5.13 recovers tasks that fail to open after frequent task switching

    An issue where some tasks could not be opened after frequent task switching, a surge in background notifications, or failure to load conversation content has been fixed.

    Usage details

    Who can use it: Qoder 0.5.13 users.

    How to use it: Update Qoder to version 0.5.13 or later; try reopening the task that will not open from the task list.

    Tip: When switching between many tasks quickly, first confirm that the active task has been saved.

    ✓ Verified by official sourceSource: Qoder sürüm notları

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 45

  • FeatureGrok Bot

    Shopping cards filter out products without real photos

    When other products have real photos, products without photos are excluded from shopping cards; cards without photos appear as a text list.

    Usage details

    Who can use it: Grok Bot shopping results.

    How to use it: Run a shopping search; review products with photos in cards and those without photos in the text list.

    Tip: If visual comparison is needed, prioritize products with real photos in the cards.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Auto-review assesses large and binary file writes based on risk

    Auto-review can now make decisions about large and binary file writes on the Bot computer; operations deemed safe proceed without showing an approval card.

    Usage details

    Who can use it: Grok Bot users who use Auto-review.

    How to use it: Turn on Auto-review in the Bot computer settings and observe the recommended decisions for large/binary file operations.

    Tip: Check Auto-review logs for production files, even if your team policy permits it.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • Plugin / IntegrationGrok Bot

    Unconnected plugins show a direct Connect action in the @ menu

    Unconnected plugins in the @ menu are now listed with only the plugin name and the Connect action.

    Usage details

    Who can use it: Grok Bot users who use plugins.

    How to use it: Open the @ menu in the chat composer; click Connect for the plugin you want to use.

    Tip: Connect the required plugin before starting a task to reduce authorization prompts along the way.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Bot confirms delivery contact details before the payment card

    Before showing the order card, Bot confirms payment contact details such as the receipt email, reducing stalled post-Pay checkout flows.

    Usage details

    Who can use it: Grok Bot connected-store checkout flows.

    How to use it: Review the email and other contact details on the order card before pressing Pay.

    Tip: Verify the receipt address before payment to prevent the invoice from going to the wrong account.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Bots can open in separate windows from the right-click menu

    A Bot can be opened in a separate window using New window in the right-click menu; double-clicking in the sidebar no longer detaches a window unintentionally.

    Usage details

    Who can use it: Grok Bot desktop app users.

    How to use it: Right-click the Bot in the sidebar and select New window.

    Tip: When working with multiple Bots, move each Bot to a separate window to avoid mixing up tasks.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    A New updates available notification appears for changes to an open Page by the Bot

    When the Bot rewrites an open Page, a New updates available badge appears, and it can load the latest version at your current scroll position.

    Usage details

    Who can use it: Grok Bot Pages users.

    How to use it: Wait for the Bot to change an open Page, then click the New updates available badge that appears.

    Tip: Use the badge to load the update without losing your place on a long Page.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    The first rows of large Excel files appear sooner

    The first rows of large Excel workbooks appear immediately while the rest of the file loads in the background.

    Usage details

    Who can use it: Users who upload large Excel files to Grok Bot.

    How to use it: Open a large Excel workbook and start reviewing it without waiting for the first rows to appear.

    Tip: After verifying the data structure from the first rows, wait for the remaining sheets to load.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Confirmation is required for Take over and End call during live calls

    Take over and End call actions on the live phone call card now require confirmation first to prevent accidental handoffs or hang-ups.

    Usage details

    Who can use it: Grok Bot phone call users.

    How to use it: Click Take over or End call on the live call card and confirm in the dialog that appears.

    Tip: Check the number and call status on the card before ending the call.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • Plugin / IntegrationGrok Bot

    Connector controls distinguish Sign In, Sign In Again, and Sign Out states

    Connector controls now clearly show the Sign In, Sign In Again, and Sign Out actions; locally running apps are labeled Runs on your computer after sign-in.

    Usage details

    Who can use it: Grok Bot Connect Apps and local connector users.

    How to use it: Open a connector in Connect Apps; use Sign In, Sign In Again, or Sign Out, depending on its status.

    Tip: After connecting a local app, verify where the data is processed by checking for the Runs on your computer label.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Project and Library show a clear status message when Context sync is off

    When Context sync is off, the Notes page in Project and Library now show the message Context sync is not running instead of a generic loading error.

    Usage details

    Who can use it: Grok Bot Project and Library users who use Context sync.

    How to use it: Turn off Context sync or stop the connection; open Project Notes or Library and check the status message.

    Tip: When you realize your notes are not updating, re-enable sync instead of making decisions based on old content.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Team admins can turn off proactive messages from Bots

    Team admins can now turn off Proactive messages for the team’s Bots from the dashboard.

    Usage details

    Who can use it: Team admins on the Team plan.

    How to use it: Open team settings in the Grok Bot dashboard and turn off Proactive messages for Bots.

    Tip: Instead of stopping proactive tasks entirely, turn them off only for team Bots that create noise.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • Plugin / IntegrationGrok Bot

    Sign in to Enterprise connectors with organization SSO

    According to the RSS changelog entry, connectors that use an organization’s single sign-on connect through the Sign in with SSO flow and return to the Bot after signing in through the browser.

    Usage details

    Who can use it: Connectors using organization SSO on the Enterprise plan.

    How to use it: Select Sign in with SSO on an Enterprise connector, then wait to return to the Bot after signing in through your organization’s browser flow.

    Tip: If the SSO redirect doesn’t complete, check the browser’s pop-up and third-party cookie permissions.

    ✓ Verified by official sourceSource: Grok Bot RSS changelog v1.2.0
  • CapabilityGrok Bot

    When there’s no match, the Bot relaxes filters and explains nearby options

    If no results are found with airline, departure time, or layover filters, the Bot shows the closest options and specifies which filters it relaxed.

    Usage details

    Who can use it: Grok Bot flight search users.

    How to use it: Run a filtered flight search; if there are no results, open the explanation of the relaxed filters shown.

    Tip: If there’s no exact match, compare the suggested nearby options and check the price before narrowing the filters again.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Expanded Mermaid diagrams can be copied as PNGs

    An option to copy the diagram as a PNG by right-clicking has been added to the expanded Mermaid diagram view.

    Usage details

    Who can use it: Grok Bot users who create Mermaid diagrams.

    How to use it: Expand the diagram, right-click it, and select Copy Image.

    Tip: Copy the diagram in expanded view before adding it to a presentation so the small chat preview doesn’t limit its resolution.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    QR verification card appears in Google phone checks

    The Google phone check in a Bot message now displays a verification card with a QR code that can be scanned or opened on a phone.

    Usage details

    Who can use it: Grok Bot 1.1.0 and later; plan information not specified.

    How to use it: Open the Google verification card in the Bot message and scan the QR code with your phone or open the link on the card on your phone.

    Tip: When pairing with a phone, scanning the QR code directly with the device’s camera is the fastest way.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Task files and colored task labels arrive on the Tasks page

    On the Tasks page, the file badge lists and opens files; task labels appear as colored badges.

    Usage details

    Who can use it: Grok Bot Tasks users.

    How to use it: Click a task’s file badge on the Tasks page, or use the colored labels to identify the task.

    Tip: Open files from the relevant task’s file badge instead of searching through chat history.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    The update flow shows progress and automatically restarts

    When Update is selected, the window closes immediately; the sidebar shows the checking, download, and preparing steps, and Grok Bot restarts when it’s ready.

    Usage details

    Who can use it: Grok Bot desktop app users.

    How to use it: Click the Update button in the sidebar and monitor the download/preparation status.

    Tip: Don’t close the app manually during the update; it will restart automatically when ready.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    HTML previews work without network access

    HTML file previews now work without network access; remote scripts, stylesheets, fonts, and images aren’t loaded.

    Usage details

    Who can use it: Grok Bot users who preview HTML files.

    How to use it: Upload an HTML file to Bot and open the preview; check that only the file’s local content is processed.

    Tip: If a page that relies on external resources looks incomplete, add the required assets locally to the file.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • Plugin / IntegrationGrok Bot

    Out-of-scope container scripts can access team secrets and setup files

    Team setup scripts configured to run outside the container now receive the team’s Team Secrets and setup files.

    Usage details

    Who can use it: Team setup scripts on the Teams plan.

    How to use it: Configure the team setup script to run outside the guest container; check the script’s access to Team Secrets and setup files.

    Tip: Give the script only the secrets it needs, and don’t print secrets in its output.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    The sidebar can be resized with the keyboard

    The sidebar resize handle can be focused with Tab; the arrow keys, Home, and End can be used to resize it.

    Usage details

    Who can use it: Grok Bot desktop app users.

    How to use it: Use Tab to focus the sidebar resize handle; adjust it with the arrow keys, or use Home/End to move to the limits.

    Tip: Use the keyboard controls for a repeatable width without using a mouse.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Command approval card shows the full command and an Allow once option

    Command approval cards now show the full command and flag hidden characters; Allow once runs the command only one time.

    Usage details

    Who can use it: Users running commands on the Grok Bot computer.

    How to use it: Open the command approval card and review the full command; select Allow once for one-time permission.

    Tip: Use Allow once instead of granting persistent permission without checking for invisible characters.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Tools disabled in desktop MCP settings are also hidden in Bot turns

    Tools disabled for a server in desktop MCP settings are now hidden and unavailable in personal Grok Bot turns and routines.

    Usage details

    Who can use it: Grok Bot users who use MCP.

    How to use it: Disable the tool on the server in desktop MCP settings, then try calling the same tool in a Bot chat or routine.

    Tip: Before running routines, check that local MCP settings haven’t disabled the tools you expect to use.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Desktop automations respond faster

    Automations that control the desktop now type and respond faster, more like interactive computer use.

    Usage details

    Who can use it: Grok Bot desktop automations.

    How to use it: Run an automation that requires desktop control; observe its typing and response steps.

    Tip: Break long automations into smaller steps to limit the impact of errors despite faster responses.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    MCP allowlist imports indicate Local/Cloud scope

    Group MCP allowlist imports show whether each rule is Local only, Cloud only, or Local and Cloud, and request approval for the first rule on each surface.

    Usage details

    Who can use it: MCP administrators on the Teams plan.

    How to use it: Start a group MCP allowlist import; review each rule’s surface label and approve the first rule.

    Tip: Verify the scope label before importing each rule to avoid mixing local and cloud surfaces.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Setup status for customer-hosted computers is now visible

    In Enterprise, the Setup column for customer-hosted computers shows the team setup script status; hovering over it reveals revision and error details.

    Usage details

    Who can use it: Customer-hosted computers on the Enterprise plan.

    How to use it: Open the customer-hosted computers list in the admin panel and hover over the status in the Setup column.

    Tip: Compare the revision in the error details with the script repository to verify which version is running.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Hotel photos open in a full-size viewer

    Hotel photos now open in a full-size viewer with the hotel name, photo counter, thumbnail previews, and previous/next buttons.

    Usage details

    Who can use it: Users viewing Grok Bot hotel search results.

    How to use it: Click a photo in a hotel result; use the thumbnail previews or previous/next buttons in the viewer.

    Tip: Before booking, review the room and property photos in sequence using the counter.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Hotel pages show ratings, websites, and map listings together

    Hotel pages show star ratings and review counts; the Website detail opens the hotel’s site, and hotels are listed next to the map.

    Usage details

    Who can use it: Grok Bot hotel search users.

    How to use it: Open a hotel page; use the review rating, Website detail, and the list of hotels next to the map.

    Tip: Compare hotels in the map list by rating instead of relying only on card order.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Hotel and car rental results use local distance units

    Search radii in hotel and car rental results are now shown in miles or kilometers based on your computer’s settings.

    Usage details

    Who can use it: Grok Bot travel search users; the unit is taken from the computer’s settings.

    How to use it: Check your computer’s region and measurement-unit settings, then search for a hotel or rental car.

    Tip: Keep in mind that the radius unit may vary when comparing results across devices.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Zoom and address copying added to hotel and place maps

    Hotel and place maps support zooming in, zooming out, and recentering by dragging; hotel pages have a Copy address option.

    Usage details

    Who can use it: Grok Bot’s travel and place search flows.

    How to use it: Use the map controls on a hotel or place page; on a hotel page, click Copy address.

    Tip: Use the copied address as a search criterion on a booking site to reduce location errors.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Project sharing clearly states who can participate

    When asked to share a Project, the Bot returns the Project page and says whether participation is limited to the team, invitees, or the owner only.

    Usage details

    Who can use it: Users who use Grok Bot Project sharing; scope depends on team settings.

    How to use it: Ask the Bot to share the Project, then check the participation scope on the returned Project page.

    Tip: For sensitive Projects, don’t share the link without verifying the invitee scope.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Quick Chat selection keeps its color highlight and macOS page context

    In Quick Chat, the selected item is highlighted in the Bot’s color; on macOS, the Bot also sees the page title, address, and selected text.

    Usage details

    Who can use it: Quick Chat; macOS context information for macOS users.

    How to use it: Open Quick Chat and select an item on the page; on macOS, verify that the page title, address, and selected text are sent.

    Tip: On long pages, narrow the Bot’s context by selecting only the text you need.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Follow-up messages preserve the context of the current task

    Follow-ups, correction requests, and retry requests sent to a tracked task now continue from the current task’s results instead of starting a new task.

    Usage details

    Who can use it: Users with tracked tasks in Grok Bot.

    How to use it: Send a follow-up or correction message in the chat for a running or completed tracked task.

    Tip: If you don’t want to start a new task, send the follow-up message in the same task thread.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    The Tasks page groups completed work in a single Closed section

    Done, Canceled, and Archived tasks are now grouped in a single Closed section with filters and search.

    Usage details

    Who can use it: All users of the Grok Bot Tasks page.

    How to use it: Open the Tasks page, go to the Closed section, and select the Done, Canceled, or Archived filter.

    Tip: Use search within Closed instead of a status filter to find an old task.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • AvailabilityGrok Bot

    Team Bot skills are read-only for users without editing permissions

    Team members who can use a Team Bot but cannot edit it can click skills in the chat and view their contents in read-only mode.

    Usage details

    Who can use it: Team members on the Teams plan who are authorized to use the Bot.

    How to use it: Click a skill name in the Team Bot chat; read the content that opens without editing permissions.

    Tip: Check the Bot’s team scope and visibility level before sharing skill instructions.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Team Bot welcome messages can identify its role and the team accounts it uses

    When you’re added to a Team Bot, the welcome message can explain the Bot’s role, owner, and the team accounts it uses.

    Usage details

    Who can use it: Team Bots on the Teams plan.

    How to use it: Get added to a Team Bot or open an existing Team Bot conversation; read the role and account information in the welcome message.

    Tip: Don’t send a sensitive request without verifying which account the Bot will use.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • SettingGrok Bot

    Team admins manage network mode and blocked sites in Security

    Team admins can now manage the Grok Bot network mode, allowed destinations, and team blocked sites in the Security section, alongside organization blocked sites.

    Usage details

    Who can use it: Team admins on the Teams plan.

    How to use it: Open the Security section in Grok Bot team settings to edit the network mode, allowed destinations, and blocked sites.

    Tip: Keep in mind that organization-level blocks may impose broader restrictions than team-level ones.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Order tracking can check hourly as delivery approaches

    Order tracking checks more frequently as delivery approaches, increasing checks to once an hour when the package is out for delivery.

    Usage details

    Who can use it: Grok Bot’s connected-store order tracking.

    How to use it: Place an order at a connected store or open an existing order card and monitor its tracking status.

    Tip: After you receive an out-for-delivery notification, ask the Bot for an updated status to get the delivery PIN or time window.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Follow-up responses in meetings are detected even when Bot isn’t addressed by name

    A response to Bot’s question or a continuation of its request within one minute is detected, even if Bot isn’t addressed by name.

    Usage details

    Who can use it: Plans that use Grok Bot’s meeting attendance feature.

    How to use it: Respond directly to a question Bot asks or a request it makes during the meeting within one minute.

    Tip: Keep your response brief and relevant to the question to avoid confusion with other conversations.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Flight searches compare more fare sources

    Flight searches now check fares from more sources, returning more reliable results for comparison.

    Usage details

    Who can use it: Grok Bot flight search users.

    How to use it: Search for a flight on the same route and date, then review the source and comparison options.

    Tip: Compare fare classes and refund policies before choosing the cheapest result.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Flight search shows its steps and hides Book while processing

    Flight searches display steps such as Searching flights and Checking the fare; the Book button is hidden while the Bot is booking.

    Usage details

    Who can use it: Grok Bot flight search users.

    How to use it: Start a flight search and follow the processing steps on the result card.

    Tip: If the Book button isn’t visible, wait for the fare check or booking process to finish.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Flight results updated with details and booking views

    When a flight is selected, Full details and Book options are shown; expanding the result displays baggage, change, and refund policies.

    Usage details

    Who can use it: Grok Bot’s flight search and booking flow.

    How to use it: Select a flight result; review the conditions with Full details, then proceed to booking with Book.

    Tip: Check the baggage and refund policies in Full details before clicking Book.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • FeatureGrok Bot

    Product pages expand options, About and Reviews sections

    Product pages show four options directly for each group and let you view the rest with Show all; they also offer About and Reviews sections with ratings and review counts.

    Usage details

    Who can use it: Grok Bot shopping flows.

    How to use it: Select Show all on the product page; review product information in the About and Reviews sections.

    Tip: Before purchasing, use Show all to check all variants and Reviews to check the rating distribution.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0
  • CapabilityGrok Bot

    Remote browser starts faster by inheriting session information

    It starts a few seconds faster when the remote browser inherits logins, cookies, and open tabs.

    Usage details

    Who can use it: Users of Grok Bot’s remote browser.

    How to use it: Open a Bot browser session that has already been logged in; wait for open tabs and session information to be inherited.

    Tip: Use a separate browser session for untrusted tasks to avoid carrying over old logins.

    ✓ Verified by official sourceSource: Grok Bot changelog v1.2.0

Watchlist

Kimi K3.1 Pro

46
Moonshot AI · Kimi

11th place; still the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. Kimi Code 2.6.3 fixed Shell output, the Windows WebBridge update, and session-resume/forking bugs; update with `kimi upgrade`.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; 12th place with 44 points. Available in preview on Google Agent Platform as `muse-wonder-1.0`.

Qwen4 Max

43
Alibaba · Qwen

Launched on November 16; 13th place with 43 points. Available in Model Studio as `qwen4-max` (1M context, Singapore pricing $1.50 / $4.50), QwenCloud Token Plan, QwenWork, and QwenStudio. Now also available on OpenRouter as `qwen/qwen4-max` (no hosting within the EU); also arrived at providers such as Novita, You.com, and Runware. Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

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

Muse Ember

36
Meta · Meta AI

Launched on November 9. Meta’s smallest Muse model; $0.10 / $0.50, 1M context, `muse-ember-1.0` on Meta Model API and included in all Muse Code plans.

Grok 4.7

35
xAI · Grok

The current model for the Grok app and API. Grok 4.8 has no official API ID, price, or Artificial Analysis score yet. Grok Bot v1.2.0 launched on November 16: SSO for Enterprise connectors, full command display and “Allow once” for command approval, and network mode settings for Team administrators.

Sources