YZ·RADARyz.mikail.net
Daily edition · 17 November 2026

Tencent Hy4 flash launches, ChatGPT opens computer-using agents to healthcare

There are no changes to the TOP 10 today either. The rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are unchanged from yesterday: Claude Fable 5.2 leads with 65 points, followed by Gemini 4 Carbon at 60, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. GPT-6.1 Sol and Claude Sonnet 5.5 are tied at 55, Gemini 4 Hydrogen has 54, GPT-6.1 Terra 51, and Gemini 4 Argon and Muse Wonder Full 50 points. The watchlist is unchanged too: Kimi K3.1 Pro is 11th with 46 points and remains the strongest open-weight model; it is 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. Tencent’s Hy4 flash, announced today, is not yet listed on Artificial Analysis. LMArena’s live page was again blocked by a CAPTCHA; the latest accessible view was dated November 11.

Healthcare products were the focus of OpenAI’s day. Computer-using agents that read screens, click, type and navigate between systems can now be used in healthcare and administrative workflows through ChatGPT for Healthcare, ChatGPT for Clinicians and the API. Access controls, human approval and audit logs are supported; for existing customers, a workspace administrator enables the capability. The clinical search and citation features in ChatGPT for Clinicians were also updated. OpenAI reports higher scores across eight dimensions in evaluations by more than 1,300 physicians. More than 20 plugins were added to the healthcare plugin directory, including Aiva, Luma Health, Notable, Redox, Stedi, TriZetto, Turquoise Health and UpToDate Expert AI. On child safety, the free `omni-moderation-csam-latest` image classifier is now available to approved organizations. ChatGPT also displays a warning when users attempt to search for CSAM, removes the conversation and pauses access for 24 hours. A partnership with LVMH’s media division will bring content from publications such as Les Echos and Le Parisien to ChatGPT in summaries with source links. For developers, Agents SDK Python v0.24.0-alpha.8 added managed approval before tool calls, resuming from a budget-preserving checkpoint, and application-specific model clients such as LiteLLM. JavaScript v0.22.0-alpha.3 added a read-only workspace and the restricted `trusted_workspace_write` setting.

Almost all of Anthropic’s 40 entries are for Claude Code v2.1.326. Code Mode can now be enabled alongside agent teams. Untrusted folders added with `--add-dir` in the local terminal now require trust approval; `claude --bg` rejects such a folder unless it has been trusted in advance. Writes by sandboxed commands to sensitive home-directory folders have been restricted. `/status` shows which features are disabled because of the provider or an environment variable. In VS Code, prompt suggestions can be turned off with `claudeCode.promptSuggestions`. A mix-up of long results in parallel MCP calls was also fixed. On Claude Platform, text watermarking has been extended to outputs from Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5; watermarking is not guaranteed if temperature, top_p or top_k is set to a non-default value. In Google, Gemini Notebook has started generating shareable interactive sites (visual explainers, timelines, dashboards) from your sources. Gemini Skills can now be shared with View or Edit permissions and can use files from Drive, Notebook or Photos as sources. The Gemini side panel in Apps Script is now generally available. Google Pics has moved to Nano Banana 2.1, and Gems has moved to the Settings panel. Usage of Gemini-powered features in Looker that exceeds the monthly token allocation will be charged starting February 1, 2027. DeepSeek-V4.1-Flash-1017 also entered preview in Model Garden.

Tencent announced Hy4 flash, with 224 billion total parameters and 12 billion active parameters. The model supports multimodal input and a 1M context window, and is positioned for everyday productivity tasks. On the TokenHub API, it costs $0,070 per 1M input tokens, $0,281 per 1M output tokens, and $0,003 per 1M cached input tokens. It is free on WorkBuddy and 50 percent off on TokenHub through November 30; open weights are planned for release later. Alibaba added open caching to the Qwen4 API, with a TTL of 5 minutes or 1 hour. It also brought MindLab’s Macaron models to Model Studio Singapore and made Qwen4-Max available in QwenWork and QwenStudio. Qwen Code v0.27.4 added `/loop 5m …` schedules for Managed Sessions; AUTO mode now requires explicit instructions for destructive actions. Meta has started testing Instagram’s standalone Direct messaging app in Türkiye and Thailand; the app can also be used with an email address or phone number, without an Instagram account. Reports say DeepSeek’s free web and app chat has gained restrictions, including a daily quota, a message limit of about 40,000 characters and the inability to continue chats left idle for a long time. These reports are based on articles by Zaobao and Sina Finance; DeepSeek has not officially confirmed them. Four fixes merged into Xiaomi’s MiMo Code main branch are not yet in a packaged release. There were no new user-facing changes on the official channels of Moonshot, xAI and Z.ai between November 15 and 17. A total of 144 new developments were added to the archive today.

Today’s TOP 10

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

  1. 01

    Claude Fable 5.2

    Anthropic · Claude=
    65pts

    Released on November 9, it leads with 65 points. Available on all paid Claude plans; `claude-fable-5-2` on 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

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    `gemini-4-carbon` on the Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, $4 / $20 from December 1). Antigravity’s new Orchestrator mode assigns planning and tough decisions to Carbon, and routine edits and tests to Hydrogen; Orchestrator (Carbon) is on Ultra.Released: 28 October 2026

  3. 03

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

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

  4. 04

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

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

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

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

  6. 06

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

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

  7. 07

    Gemini 4 Hydrogen

    Google · Gemini=
    54pts

    Released on November 12; scores 54 with the xhigh setting, at a cost of $3,16 per task ($5,48 for Carbon). `gemini-4-hydrogen` is also available on the free API tier; paid pricing is $1,50 / $7,50. The top model in the Gemini app for Google AI Plus subscribers. Since November 16, it can also be selected 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 in ChatGPT Work, Codex and the API; focused on vision and computer use. The management preview for Dots agents runs through November 18.Released: 26 October 2026

  9. 09

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available on Google AI Pro/Ultra and the paid API; $2 / $10 promotional pricing 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. The Muse app added a feature this week for paid subscribers in the US to call businesses, along with the Muse from Meta Chrome extension.Released: 12 October 2026

Key signals

  • The TOP 10 and watchlist are unchanged: Claude Fable 5.2 (65) leads, and Kimi K3.1 Pro (46) is the strongest open-weight model.
  • Tencent: Hy4 flash launches. 1M context, $0,070 / $0,281 on TokenHub; free on WorkBuddy through November 30.
  • OpenAI: Computer-using agents in ChatGPT for Healthcare and more than 20 new healthcare plugins.
  • Google: Gemini Notebook creates shareable interactive sites from sources; Gemini Skills can be shared.
  • Anthropic: With Claude Code v2.1.326, Code Mode works with agent teams; text watermarking has come to older Claude models.
  • Meta’s Direct app is in testing in Türkiye. 144 new developments: Anthropic 40, Alibaba 32, OpenAI 29, Tencent 21, Google 11, Meta 4, Xiaomi 4, DeepSeek 3.

Worth trying today

01

Connect your Gemini Skills to live documents and share them with your team

Open the Skills page in the Gemini web app and create or upload a skill. When adding files, choose Google Drive, Gemini Notebook or Google Photos; the limit for uploaded files is 100 MB per skill. When a Drive document or Notebook is updated, the skill uses the latest version, so you don’t need to re-upload frequently changing company guides. To share a skill, select it, click Share and grant View access first; grant Edit access only to the person responsible for maintenance. Recipients are notified by email and can call the skill in a chat using `@` once they’ve installed it. If you want an interactive summary instead of a presentation from the same sources, type `make me a site` in a Gemini Notebook chat, describe what you want, then open and share the result with Open in a new tab.

Google · Gemini
02

Reduce costs on Qwen4 requests with open caching

Put content that stays the same across requests, such as the system prompt, tool definitions and long documents, at the start of the input; the prefix to be cached must be at least 1.024 tokens. In Model Studio’s Responses API, set mode and ttl in the `prompt_cache_options` field (5 minutes or 1 hour). Add `prompt_cache_breakpoint: {"mode": "explicit"}` to the input block where the cache should end, then send the variable question after that point. The Chat Completions and Anthropic-compatible endpoints have corresponding cache fields. For `qwen4-max`, reading from the cache costs $0,075 per 1M tokens, one-twentieth of the regular input price. Track the hit rate in the response usage fields and choose the TTL based on your workflow frequency.

Alibaba · Qwen
03

Update Claude Code to 2.1.326 and prepare for new security prompts

Run `claude update` to update and verify 2.1.326 with `claude --version`. Then run `/status` in a session; the new line shows which features are disabled by the provider or an environment variable such as `DISABLE_TELEMETRY`. When you add another folder with `--add-dir`, you’ll now be asked to approve it; inspect the `.claude` folder and hooks before approving third-party repositories. If you plan to run in the background with `claude --bg`, trust the folder in advance. Specify the folders sandbox commands need to write to individually with `sandbox.filesystem.allowWrite`. If you want Claude outputs to remain watermarked in the API, don’t send the temperature, top_p or top_k parameters.

Anthropic · Claude
04

Try Hy4 flash for free through November 30

Select Hy4 flash from the model picker in Tencent’s WorkBuddy app; it’s free there through November 30. To try it through the API, select Hy4 flash on Tencent Cloud TokenHub. The list price is $0,070 per 1M input tokens and $0,281 per 1M output tokens, with a 50 percent discount through that date. Tencent recommends this model for everyday tasks such as search, documents, slides and simple front ends; use Hy4 for long software engineering tasks. Compare the same set of prompts with your current low-cost model and keep fixed prefixes to benefit from the cached input price ($0,003).

Tencent · Hy

This edition’s updates by AI

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

Anthropic · Claude

#1 Claude Fable 5.2#3 Claude Opus 5.5#6 Claude Sonnet 5.5Claude Haiku 5.5
Full history (836) →

New in this edition · 40

  • AgentClaude Code

    Agent plan approval and the claude-api default have been updated

    The subagent plan approval card now has a two-option keyboard flow. Unless you choose to host your own agent loop, the claude-api skill now creates agents on Managed Agents by default.

    Usage details

    Who can use it: Claude Code agent view and the claude-api skill; 2.1.326 and later.

    How to use it: Use the arrow keys and Enter on the plan approval card; explicitly select this in the skill flow if you want to host your own loop.

    Tip: Check storage, provider, and history requirements before using Managed Agents.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AgentClaude Code

    Code Mode can now be enabled alongside agent teams

    Code Mode can now be enabled in sessions where agent teams are active. If Code Mode was already enabled, new sessions with teams start directly in this mode.

    Usage details

    Who can use it: Claude Code 2.1.326 and later; sessions using Code Mode and agent teams.

    How to use it: Enable Code Mode with the relevant Claude Code setting, then enable agent teams.

    Tip: Test permission and sandbox policies before using Code Mode for team tasks.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • SettingClaude Code

    Auto mode selectively applies shell permissions with wildcards

    Auto mode ignores some wildcard shell permissions that do not specify a program or begin with $HOME. If the managed autoMode.environment list does not include $defaults, it replaces the built-in entries.

    Usage details

    Who can use it: Claude Code and Claude Desktop/Agent SDK Auto mode configurations.

    How to use it: Specify the program in wildcard permissions and add $defaults to the managed autoMode.environment list.

    Tip: Rewrite broad wildcard rules on a per-command and per-target-directory basis.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • SettingClaude Code

    Auto mode defaults are more consistent until policy loads

    In organizations with HIPAA configurations, terminal and VS Code sessions can start in Auto mode by default. If no mode is specified, Auto mode starts instead of Manual until the policy loads; the server then applies defaultMode.

    Usage details

    Who can use it: Claude Code and VS Code sessions; especially Claude Code + Cowork HIPAA organizations.

    How to use it: To disable Auto mode, use disableAutoMode in managed settings; verify the mode with /status.

    Tip: In enterprise automation, check the state after the policy loads rather than the local default.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Compaction and API retries are more resilient for large inputs

    Compaction has been fixed in Bedrock, Vertex, Foundry, and telemetry-disabled sessions; image-heavy Vertex sessions can compact and retry large requests. If max_tokens exceeds the model limit, the API-reported maximum is used.

    Usage details

    Who can use it: Claude Code’s Bedrock, Vertex, Foundry, and Anthropic API providers.

    How to use it: Use /compact or automatic compaction in large sessions; the new version retries limit overruns.

    Tip: Split image and PDF uploads into smaller batches to reduce wait times and costs.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Code Mode skill and mode tool calls work in more cases

    Under Auto mode, skills and slash commands that require approval for inline commands can now prompt the user. In Code Mode, background subagents can preload skills, and modes can use Bash, Read, and other tools with tool.call.

    Usage details

    Who can use it: Claude Code Code Mode, Auto mode, skill/mode system, and background subagents.

    How to use it: Call the skill; if an inline command approval prompt appears, approve it and explicitly manage mode tool permissions.

    Tip: For third-party skills, restrict inline shell commands on a per-command basis.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Code

    Identity and background session transitions are more tightly protected

    Messages sent from another machine to sessions using an API key, auth token, profile, or workload identity may be rejected. Moving --bare sessions to the background is also blocked if there is no service key.

    Usage details

    Who can use it: Claude Code Remote Control, background, and cross-machine messaging flows.

    How to use it: Keep the same identity type in remote sessions and verify the required key in the background service environment.

    Tip: Use an explicit service identity in production services instead of a key left in the shell.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Pricing / PlanClaude Code

    Usage, budget, and renderer screens provide more accurate information

    Fullscreen /usage shows plan limits before session costs. Mode calls are included in cost and token metrics; issues with usage credits and TUI screens have been fixed.

    Usage details

    Who can use it: Claude Code paid plans, usage credits, /usage, /cost, and fullscreen TUI.

    How to use it: Check plan limits separately with /usage and session costs with /cost.

    Tip: Include model calls from modes when setting budget alerts.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    MCP tool results, schemas, and concurrent calls fixed

    Long results from parallel MCP calls no longer get mixed up; incorrect server selection for tools requested by full name has been reduced. All branches in anyOf and oneOf schemas are passed to Claude.

    Usage details

    Who can use it: Local, remote, and cloud MCP servers, as well as SDK hosts.

    How to use it: Call the MCP tool by its full name and explicitly define the type of each branch in union schemas.

    Tip: For concurrent calls, match results to requests by request ID.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Sandbox more strictly limits sensitive write paths

    Writes by sandboxed commands to sensitive home directories have been restricted on macOS, Linux, and WSL; unexpected entries inside .git are isolated. Rewritten commands also remain inside the sandbox if the initial check was sandboxed.

    Usage details

    Who can use it: macOS, Linux, and WSL sessions using the Claude Code sandbox.

    How to use it: Explicitly define the required write paths with sandbox.filesystem.allowWrite.

    Tip: Add only task directories to the allowlist instead of granting broad write access to home directories.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • SettingClaude Code

    Additional safety prompts added for worktrees, PowerShell, and bypass permissions

    Isolated worktree subagents now reject Git commands that could modify shared branches; ExitWorktree no longer deletes a worktree while work is running. More PowerShell commands now trigger a prompt or rejection flow.

    Usage details

    Who can use it: Claude Code worktree, PowerShell, and bypass permissions sessions.

    How to use it: Use ExitWorktree after completing running tasks; define required PowerShell commands with an explicit permissions.allow rule.

    Tip: Use narrow permission rules for each worktree instead of relying on bypass mode.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Code

    Trust approval is now required for new add-dir directories

    If an untrusted directory is added with add-dir or /add-dir in a local terminal, a directory trust prompt appears. claude --bg rejects such a directory unless it has been trusted in advance.

    Usage details

    Who can use it: Claude Code local terminal and self-hosted runner sessions.

    How to use it: Add the directory with --add-dir and approve it at the trust prompt; for background use, trust the directory in advance.

    Tip: Do not trust third-party repositories without inspecting their .claude contents and hooks.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AgentClaude Code

    Agent view and worktree startup flow are now safer

    In agent view, exiting an empty conversation immediately ends the session; await-initialize is rejected in incompatible background and tmux worktree combinations. Worktree and IDE checkout operations run with a more restricted PATH.

    Usage details

    Who can use it: Claude Code agent view, worktree, teleport, IDE checkout, and await-initialize flows.

    How to use it: Do not use await-initialize with worktree or tmux; provide hook and Git helper paths as absolute paths.

    Tip: Do not leave worktree hooks dependent on the shell PATH.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AgentClaude Code

    Background agent status and reporting view fixed

    Agent rows more accurately show completion, waiting, and error states; /tasks names the subagent behind a background command. Workflow resume reports the number of reused results.

    Usage details

    Who can use it: Claude Code agent view, background agents, and workflow resume.

    How to use it: Open the agent rows or the /tasks screen and check the reason for waiting.

    Tip: For multi-agent tasks, rely on /tasks status instead of the spinner.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • SettingClaude Code

    Malformed rules in CLI tool permissions now trigger a warning

    If allowedTools, disallowedTools, or tools values contain entries that do not look like complete rules, Claude Code displays a warning at startup.

    Usage details

    Who can use it: Claude Code CLI 2.1.326 and later.

    How to use it: Start Claude Code with these flags and, if a warning appears, correct each entry to use the complete rule format.

    Tip: Keep permission lists in version control and send first-run warnings to CI logs.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Platform

    Text watermarking added to supported legacy models

    Outputs from Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 carry Anthropic’s machine-readable text watermark on supported platforms. The watermark is not guaranteed if temperature, top_p, or top_k is set to a non-default value.

    Usage details

    Who can use it: Worldwide on Claude Platform API and supported Claude surfaces, as well as partner platforms such as AWS, Google Cloud, and Microsoft Foundry.

    How to use it: To preserve the watermark, do not send temperature, top_p, or top_k parameters in requests; check that your SDK does not add them automatically.

    Tip: Verify C2PA credentials and text markers in generated files with Claude Content Checker.

    ✓ Verified by official sourceModels: Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, Claude Haiku 4.5Source: Claude Platform sürüm notlarıSource: Claude’un AI üretimi içeriği nasıl işaretlediğiSource: Text watermark teknik belgesi
  • Plugin / IntegrationClaude Code

    MCP reconnection and task questions are clearer

    While MCP is reconnecting, pressing Esc or Ctrl+C twice closes the wait while allowing the connection to continue in the background. MCP task questions can be answered later, and /mcp shows the reason startup failed.

    Usage details

    Who can use it: Claude Code MCP servers and background MCP tasks.

    How to use it: Press Esc on a pending MCP question and answer it later through /tasks.

    Tip: For tasks requiring a human response, use Esc to defer; handle disconnection separately.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    Memory, Claude in Chrome, and Web Search workflows fixed

    Listing is safer in memory directories containing symlink loops or named pipes. Claude in Chrome increases the timeout from 90 to 150 seconds for long browser_batch calls; Web Search no longer shows a blank screen when retrying.

    Usage details

    Who can use it: Claude Code Memory, Claude in Chrome, and Web Search tools.

    How to use it: Open the Memory directory with /memory; retry Chrome and Web Search calls in the new version.

    Tip: Do not place named pipes or symlinks that follow other links in the Memory directory.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • FeatureClaude Code

    Model defaults and built-in command help are clearer

    /model shows the saved model alongside the plan’s actual default and really returns to the default when selected. Support for command-specific /help command has been added.

    Usage details

    Who can use it: Claude Code terminal interface 2.1.326 and later.

    How to use it: Select the default with /model; run /help command for a specific command.

    Tip: After changing models, verify the name and provider with /status.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • APIClaude Code

    Optional host identifiers can be added to OpenTelemetry

    The OTEL_RESOURCE_INCLUDE_HOST environment variable has been added. When enabled, host.name and host.id are added to resource attributes; it is off by default.

    Usage details

    Who can use it: Claude Code 2.1.326 and later; enterprise deployments using OpenTelemetry.

    How to use it: Set OTEL_RESOURCE_INCLUDE_HOST=1 before starting Claude Code.

    Tip: Check your telemetry policy to see whether host identifiers are considered sensitive data.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    The behavior of managed-mcp.json added during a session is now explained

    If an administrator adds managed-mcp.json during a session, the reason MCP servers do not load immediately is explained. The file may be applied at the next check; open sessions must be restarted for the changes to take full effect.

    Usage details

    Who can use it: Organization-managed Claude Code sessions and managed MCP configurations.

    How to use it: Review the notice and restart Claude Code for all changes to take effect.

    Tip: Apply organization policy changes at the start of a new session, not in the middle of a live task.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Automatic compaction in plan mode uses the responding model

    Automatic context compaction in plan mode now uses the model responding to the user. Compaction costs may increase for plans running with Opus.

    Usage details

    Who can use it: Claude Code plan mode and automatic compaction sessions.

    How to use it: Check the selected model in plan mode with /model; track costs with /cost or /usage.

    Tip: To reduce compaction costs on long plans, clear unnecessary context at the start of the task.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    Navigation in plugin marketplaces and lists has been improved

    Support for narrow terminals, Home, End, pagination, mouse, and clicking has been improved in the Plugin Discover, marketplace, Installed, Errors, and Stats screens. If ALL_PROXY does not respond, the network request times out after 25 seconds.

    Usage details

    Who can use it: Claude Code plugin and marketplace interface.

    How to use it: Use the Home, End, and pagination keys on the /plugin screen; test CONNECT and timeout behavior for your proxy.

    Tip: Verify marketplace access through your corporate proxy using both CONNECT and timeout.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    Plugin MCP bundles are cleaned up during removal

    The .mcpb and .dxt files extracted by a plugin are also deleted when the plugin is removed; a shared bundle is kept if another plugin uses it.

    Usage details

    Who can use it: Claude Code plugin and worktree users; 2.1.326 and later.

    How to use it: Remove the plugin through /plugin and check whether shared bundles are preserved.

    Tip: Inventory shared MCP bundles before removing a plugin.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Code Remote Control

    Remote Control proxy connections and debug logs have changed

    Remote Control uses a CONNECT tunnel through the proxy; a proxy that rejects CONNECT stops the connection with an error. verbose debug logs have moved to ~/.claude/child-logs/.

    Usage details

    Who can use it: Claude Code Remote Control behind a proxy; 2.1.326 and later.

    How to use it: Allow CONNECT tunneling on the proxy; inspect detailed logs using verbose and ~/.claude/child-logs/.

    Tip: Also test CONNECT permission for the target host on your organization’s proxy.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Code

    Remote, cloud, and Artifact file flows are more reliable

    File uploads in Remote Control and cloud sessions are retried in response to some proxy 5xx responses. Artifact publishing and external sharing permissions are handled more accurately.

    Usage details

    Who can use it: Remote Control, cloud/self-hosted runner, Claude.ai, Cowork, and Artifact.

    How to use it: If file delivery fails, wait for a retry; check the sharing scope when publishing an Artifact.

    Tip: Do not automate sharing without reading the send confirmation after the secret-key scan.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code self-hosted runner

    Built-in gh api path added for self-hosted runners

    On Linux and macOS self-hosted runners, if gh is unavailable, the built-in gh api through Anthropic is used by default and can use the existing GitHub login. It can be disabled with --use-anthropic-gh false.

    Usage details

    Who can use it: Linux and macOS self-hosted runners; sessions without gh installed.

    How to use it: Start the runner normally or pass --use-anthropic-gh false; check GitHub authentication behavior.

    Tip: If required by your data boundary, disable the built-in path and provide your own gh installation.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • FeatureClaude Code

    status shows which features are disabled

    A new line has been added to the status output indicating which features are disabled by the provider or environment variables such as DISABLE_TELEMETRY.

    Usage details

    Who can use it: Claude Code 2.1.326 and later.

    How to use it: Run /status in the session and inspect the provider or environment variable disabling features.

    Tip: When a feature disappears, use /status to distinguish policy from environment variables before changing settings.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Terminal streaming, retries, and prompt editing are more robust

    Issues causing transcript lines to disappear during streaming, long queues to obscure responses, and fullscreen freezes have been fixed. Paste, multiline prompt, undo, and question-answer flows also work more reliably.

    Usage details

    Who can use it: Claude Code terminal and fullscreen interface.

    How to use it: View long responses in fullscreen and verify pasted content before sending it.

    Tip: Confirm critical results in the transcript after streaming is complete.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Terminal status protocol and accessibility indicators improved

    Terminals that support OSC 7501 report background commands, monitor, and MCP tasks more accurately. The progress bar also displays waiting and suspended states correctly.

    Usage details

    Who can use it: Terminals that support OSC 7501 and sessions using a screen reader.

    How to use it: Restart the terminal and test OSC 7501 statuses and /config accessibility settings.

    Tip: Verify waiting, background, and user-response states separately.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Responsiveness and resource usage improved in long sessions

    Responses after file operations, streaming, and resize responsiveness have improved; idle CPU usage has been reduced. Closed transcript token counting is less frequent for multiple subagents, and the first Bash command is faster on Windows.

    Usage details

    Who can use it: Claude Code terminal, Windows, Claude Desktop, and multi-agent sessions.

    How to use it: Update to 2.1.326 and remeasure long sessions using the same model and repository.

    Tip: Stop subagents when the work is done; this also reduces costs.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • AvailabilityClaude Code for VS Code

    VS Code sign-in and IDE connection are more reliable

    Fixed issues where saved sign-in in VS Code was lost on some networks, the keyless Console sign-in screen appeared unnecessarily, and a pasted verification code disrupted the session. If a second session takes over the connection, instructions for returning to the first session are shown.

    Usage details

    Who can use it: Claude Code for VS Code 2.1.326 and later.

    How to use it: Sign in from the VS Code panel; if a second session takes over the connection, follow the displayed instructions for returning.

    Tip: In multiple workspaces, check which Claude Code session the panel is connected to.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • FeatureClaude Code for VS Code

    VS Code sandbox, live output, and Artifact links improved

    Fixed an issue where the VS Code sandbox dialog unnecessarily created empty files; live output from running commands and the Open artifact link when an Artifact is published are now displayed more consistently.

    Usage details

    Who can use it: Claude Code for VS Code 2.1.326 and later.

    How to use it: Open the running command card in VS Code chat; use the Open artifact link when an Artifact is published.

    Tip: Treat live output as a progress signal and verify the final result in the command card.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • SettingClaude Code for VS Code

    Setting added to disable prompt suggestions in VS Code

    The claudeCode.promptSuggestions setting was added to the VS Code extension and can disable prompt suggestions.

    Usage details

    Who can use it: Claude Code for VS Code extension 2.1.326 and later.

    How to use it: Set claudeCode.promptSuggestions to false in VS Code settings.

    Tip: Turn suggestions off in intensive automation sessions and leave them on in interactive sessions.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • APIClaude Code

    The scope of managed HTTP hook and header settings is now clearer

    allowedHttpHookUrls and httpHookAllowedEnvVars are applied more specifically to managed hooks. Content-Length and Transfer-Encoding in HTTP hook headers are ignored.

    Usage details

    Who can use it: Claude Code 2.1.326 and later; organizations using managed HTTP hooks.

    How to use it: Explicitly define allowed URL and environment variable lists; generate request size from the body.

    Tip: Base hook security on allowed URL and environment lists, not on header fields.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • CapabilityClaude Code

    Hints for failed prompts and stopping subagents are now visible

    The empty prompt field displays a hint explaining what to do after a request fails. The subagent view now indicates that Esc will stop the agent.

    Usage details

    Who can use it: Claude Code 2.1.326 and later; terminal subagent view.

    How to use it: After a failed request, follow the guidance in the prompt field; press Esc to stop a running subagent.

    Tip: For long-running team tasks, confirm in the status line that Esc stops the selected subagent.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • APIClaude Code

    Empty Cowork cloud sessions no longer send telemetry events

    In Cowork cloud sessions, organization OpenTelemetry export is delayed until the first message. No session.count event is sent for sessions that receive no messages.

    Usage details

    Who can use it: Cowork cloud sessions and organization OpenTelemetry export.

    How to use it: Interpret session counts in telemetry dashboards with the first-message condition in mind.

    Tip: Track session count and active task count as separate metrics.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • Plugin / IntegrationClaude Code

    Malformed plugin.json fields are now reported visibly

    /plugin and claude plugin list indicate when malformed plugin.json fields, such as author or homepage, are ignored. The plugin can still be loaded, but administrators can see the issue.

    Usage details

    Who can use it: Claude Code plugin users and administrators; 2.1.326 and later.

    How to use it: Open the output of /plugin or claude plugin list; if there is a warning, fix the plugin.json fields and reload.

    Tip: Run claude plugin validate before deployment.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • FeatureClaude Code

    A Ctrl+O shortcut is now available to show hidden startup notices

    A Ctrl+O shortcut has been added to view startup notices that were shown and then hidden.

    Usage details

    Who can use it: Claude Code terminal interface 2.1.326 and later.

    How to use it: Press Ctrl+O in a session to show hidden startup notices.

    Tip: Use this shortcut to check for missed warnings when troubleshooting permission, policy, or sign-in issues.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları
  • APIClaude Code

    Overflow and line-limit properties added to mod interfaces

    Mod developers can use overflowX=scroll for Box and maxLines and wrap=clip for Text.

    Usage details

    Who can use it: Claude Code mod developers and mods compatible with 2.1.326.

    How to use it: Use these properties in mod components and test the layout in a narrow terminal.

    Tip: Add a short summary or continuation indicator with maxLines.

    ✓ Verified by official sourceSource: Claude Code v2.1.326 sürüm notları

Google · Gemini

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

New in this edition · 11

  • AvailabilityGemini

    Gems moved to the Settings panel in Gemini

    Gems has moved to the Settings panel in the Gemini app; users can still create, edit, and use Gems at this stage. This is a newly dated implementation step in the previously announced transition from Gems to Skills.

    Usage details

    Who can use it: Users covered by the personal and Workspace transition schedules in the Gemini app; as of November 17, 2026, Gems can be accessed through the Settings panel.

    How to use it: Open the Gemini app, go to the Settings panel, and manage existing Gems in the Gems section; use the Skills page for new workflows as well.

    Tip: Re-create important Gems as Skills now and keep their information files separately so you won’t depend on future transition steps.

  • FeatureGoogle Apps Script

    Gemini side panel in Apps Script reaches general availability

    The Gemini side panel in the Apps Script IDE is now GA; users can create Workspace automations using natural language and accept or reject code suggestions block by block. The Fix errors button in Execution Log explains runtime errors and suggests actionable fixes.

    Usage details

    Who can use it: Business Standard/Plus, Enterprise Standard/Plus, AI Expanded Access, Google AI Pro for Education, and consumer Google AI Pro/Ultra plans; phased rollout began on November 12.

    How to use it: Open Extensions > Apps Script in script.google.com or Sheets/Docs/Slides; select Ask Gemini in the top right and enter your request. For errors, use Execution Log > Fix errors.

    Tip: Review Gemini’s suggested code block by block before saving, and limit OAuth permissions to the minimum required.

  • FeatureGemini

    Gemini skills can now be shared

    You can share any skill with others from the Skills page in the Gemini app and give recipients View or Edit access. Recipients are notified by email. After reviewing and installing the skill, Gemini can use it automatically, or it can be invoked in a chat with `@`.

    Usage details

    Who can use it: Google Workspace customers and personal Google accounts; skill sharing and access management are currently available in the Gemini web app.

    How to use it: Open the Skills page in the Gemini web app, select a skill, click Share, and give the recipient View or Edit access.

    Tip: For organizational guidelines, share with View access first; allow editing only for people who own or maintain the skill.

  • FeatureGemini Notebook

    Gemini Notebook creates shareable interactive sites from sources

    Gemini Notebook can now create interactive sites from your sources; examples include visual explainers, timelines, data dashboards, and project trackers. You can open the generated result in a new tab and share it with others.

    Usage details

    Who can use it: Gemini Notebook on the web globally for users aged 18 and over, including Google Workspace subscribers. Mobile app support is coming soon.

    How to use it: Type `make me a site` in a Gemini Notebook chat and describe the site you want to create; when the result is ready, select `Open in a new tab`.

    Tip: For complex research, first organize your sources in a single notebook, then clearly describe the views and interactions you expect from the site.

  • Plugin / IntegrationGemini

    Gemini Skills can use files from Drive, Notebook, and Photos as sources

    Files from Google Drive, Gemini Notebook, and Google Photos can be added to a manually created or uploaded Gemini Skill. When Drive files and Notebooks are updated, the Skill uses the latest version; the upload limit is 100 MB per Skill.

    Usage details

    Who can use it: Google Workspace customers and personal Google accounts; in the Gemini web app.

    How to use it: Create or upload a Skill in Gemini, select Google Drive, Gemini Notebook, or Google Photos as the source when adding files, and save the Skill.

    Tip: Use a Drive or Notebook source for frequently updated guides so you don’t have to upload them again each time a document changes.

  • SettingGemini Enterprise

    Antigravity teamwork setting now also controls the `/swarm` command

    The `/teamwork-preview command` setting in AI developer tools has been renamed `/teamwork command`. In supported Antigravity versions, the same setting also enables or disables `/swarm` team agents that work with a shared message board.

    Usage details

    Who can use it: Gemini Enterprise AI developer tools and supported Google Antigravity versions.

    How to use it: In Google Cloud Console, enable or disable the `/teamwork command` setting under Gemini Enterprise AI developer tools settings > Features; then use `/teamwork` or `/swarm` in Antigravity.

    Tip: To enable multi-agent work in a controlled way across your organization, first grant `/teamwork command` access only to a pilot group.

  • OtherCodeMender

    CodeMender v0.22.0 adds time limits, CSV, and new exit codes

    CodeMender v0.22.0 adds `--duration`/`--timeout` and `CM_DURATION` support for `cm find`, `cm fix`, `cm verify`, and `cm session resume`. `cm find` and `cm report` can export and import CSV; exit code 2 is used for general errors and prompt-injection stops, while 1 is used for CI gate violations.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform users and CodeMender v0.22.0.

    How to use it: Add `--duration` or `--timeout` to commands, or set the `CM_DURATION` environment variable; for CSV, use `cm find --format csv`/`cm report --format csv` and `cm report import`.

    Tip: In CI, handle 1 separately as a policy gate violation and 2 as a general error; set a time limit for long scans.

  • ModelGemini Enterprise Agent Platform

    DeepSeek-V4.1-Flash-1017 enters Public Preview in Model Garden

    DeepSeek-V4.1-Flash-1017 has launched in Public Preview as a fully managed model by Google in Gemini Enterprise Agent Platform Model Garden. The update improves reasoning, coding, agentic, and visual performance, and supports context of up to 1 million tokens with text and image input.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform Model Garden; a MaaS model offered in Public Preview through a global endpoint.

    How to use it: Open the `DeepSeek-V4.1-Flash-1017` model card in Model Garden, enable preview access, and send a request through the managed model endpoint.

    Tip: When using a 1-million-token context, remove unnecessary history and limit the long context to only the sources you actually need.

    ✓ Verified by official sourceModels: DeepSeek-V4.1-Flash-1017Source: Google Cloud release notes — November 16, 2026
  • AvailabilityGemini Enterprise

    Gemini 4 Hydrogen becomes available in Gemini Enterprise AlphaEvolve

    Gemini Enterprise AlphaEvolve experiments can now generate candidate programs with Gemini 4 Hydrogen. The model does not replace the default; it must be explicitly selected in generation settings.

    Usage details

    Who can use it: Gemini Enterprise AlphaEvolve; global, us, and eu regions.

    How to use it: Open the AlphaEvolve Generation settings section and specify `gemini-4-hydrogen` as the model.

    Tip: Before selecting Hydrogen for candidate program generation, compare results with the default model to measure differences in speed and quality.

    ✓ Verified by official sourceModels: Gemini 4 HydrogenSource: Google Cloud release notes — November 16, 2026
  • CapabilityGoogle Pics

    Google Pics upgraded to Nano Banana 2.1

    Google Pics has moved to Nano Banana 2.1 for its existing image generation and editing workflows. The update promises higher image quality, more faithful edits, sharper text generation, and speed similar to the previous model.

    Usage details

    Who can use it: Users of Google Pics on Business Standard/Plus, Enterprise Standard/Plus, Google AI Plus/Pro/Ultra, Google AI Pro for Education, and AI Expanded Access; full rollout in 1–3 days starting November 16.

    How to use it: Open Google Pics or launch Pics in Google Docs/Slides; use the new generation and editing workflows without selecting a separate model.

    Tip: To reduce unintended changes to the surroundings, clearly specify which objects, background, and brand elements to preserve in your editing prompt.

  • Pricing / PlanLooker

    Usage overages for Gemini-assisted features in Looker will be billed

    Google announced that billing will begin on February 1, 2027, for usage exceeding the included monthly data token allocation for Gemini-assisted features generally available in Looker. Preview features are not billed; usage is measured in input and output data tokens.

    Usage details

    Who can use it: Looker customers; GA Gemini-assisted features such as Conversational Analytics. Billing begins on 2027-02-01.

    How to use it: As a Looker administrator, go to Admin > Gemini in Looker and connect the Google Cloud project to associate with the sample, with a valid billing account; then monitor token usage in the Conversational Analytics System Activity dashboard.

    Tip: Regularly monitor input/output token trends and reduce long-context queries to see your monthly allocation and potential overage charges in advance.

OpenAI · ChatGPT

#4 GPT-6 Astra#5 GPT-6.1 Sol#8 GPT-6.1 Terra
Full history (339) →

New in this edition · 29

  • CapabilityChatGPT

    Warnings and 24-hour access pauses introduced for CSAM searches

    ChatGPT now displays new product warnings for attempts to sexualize children or search for CSAM, removes the conversation, and pauses access for 24 hours. The warning explains the harm caused by the behavior and its potential legal consequences; repeated violations can lead to longer restrictions, account closure, or reports to authorities.

    Usage details

    Who can use it: In the ChatGPT product; OpenAI's announcement did not specify separate restrictions by plan or country. Referrals to specialist support are included in the warning flow.

    How to use it: No setting needs to be enabled; the warning and pause flow is applied automatically when an attempt is detected as a policy violation.

    Tip: Do not test safety features with harmful content in a real user account; product teams should use OpenAI's CSAM guidance and reporting processes.

  • Plugin / IntegrationOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK expanded native tools and MCP controls

    Native tools now offer namespace, concurrency limits, caller constraints, managed approval, client-side tool search, and error-format options. Native MCP adds discovery filters, cache controls, per-tool approval, and alias support.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; JavaScript/TypeScript applications using native tools and MCP.

    How to use it: Set namespace, maxFunctionConcurrency, approval, discovery filter, and cache options in tool and MCP configuration; enable tool search only when needed.

    Tip: Review the default concurrency acceptance behavior for side-effecting callbacks; use maxFunctionConcurrency: 1 if needed.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • SettingOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK added read-only workspaces and a trusted write setting

    Local executors support read-only workspaces and optional trusted_workspace_write configuration on Linux and macOS. Trusted writes allow broader host reads while limiting writes to specified roots; applications can also enable AGENTS.md discovery.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; Linux and macOS local executors.

    How to use it: Use a read-only workspace in local execution configuration; when controlled writes are needed, explicitly enable trusted_workspace_write and define writable roots.

    Tip: Don't allow broad host reads by default; limit writable roots to the project directory.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • ModelOpenAI API

    Free CSAM Image Classifier model made available to approved organizations

    OpenAI released omni-moderation-csam-latest, a model that detects potential child sexual abuse material and potential AI-generated CSAM using separate signals. The model helps prioritize images for human safety teams to review and is available free of charge to approved organizations.

    Usage details

    Who can use it: Vetted/approved child-safety organizations; access requires an application and OpenAI approval. The model classifies an image; it does not automatically report it to authorities.

    How to use it: Using an approved organization API key, send a POST /v1/moderations request, set the model field to omni-moderation-csam-latest, and use a single image_url in input that is no larger than 20 MB; do not submit known CSAM.

    Tip: Use category_scores to rank a human review queue rather than to make automatic accusations; they are not calibrated probabilities.

  • AgentOpenAI Agents SDK for Python

    Python Agents SDK adds managed approval before tool calls

    Agents SDK v0.24.0-alpha.8 adds a flow for reviewing and approving or rejecting arguments before Python tool calls run. The decision can be remembered for the rest of the Session; options to run input policy before approval and complete a tool batch with FinishTurn are also available.

    Usage details

    Who can use it: openai-agents-python v0.24.0-alpha.8 preview; applications using the Python native tool loop.

    How to use it: Install the alpha version, define a managed approval callback, and enable the run_input_policy_before_approval=True setting when needed.

    Tip: For tools with side effects, request approval after showing an argument summary, the target resource, and the expected effect.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • CapabilityOpenAI Agents SDK for Python

    Python Agents SDK can resume pending work while preserving budgets

    Alpha.8 saves supported pending tool requests and approvals to a checkpoint and resumes them. Previously spent request and token budgets are preserved on restore; storage options owned by the application also expand with SQLAlchemySQLiteSessionStore and CustomStorage.

    Usage details

    Who can use it: v0.24.0-alpha.8 preview; supported checkpoint, Session, and storage flows.

    How to use it: Save the checkpoint, restore it with alpha.8, and apply the request/token budget settings from the Restore a checkpoint with a budget example.

    Tip: Back up saved state before upgrading to the alpha; older alpha versions may not be able to open the new checkpoint format with budgets.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • Plugin / IntegrationChatGPT for Healthcare

    More than 20 new plugins added to the healthcare plugin ecosystem

    More than 20 new healthcare plugins used in clinical care, care coordination, and business operations have been added to ChatGPT for Healthcare. Examples include Aiva, Luma Health, Notable, Redox, Stedi, TriZetto, Turquoise Health, and UpToDate Expert AI; plugins can be used together in the same conversation.

    Usage details

    Who can use it: ChatGPT for Healthcare customers; selected plugins are also available to ChatGPT for Clinicians users. Workspace and administrator permissions apply in HIPAA-enabled Enterprise workspaces.

    How to use it: Ask your workspace administrator to enable the required plugins, install the relevant plugin from the Plugin Directory, connect the provider account, and explicitly mention the plugin name or workflow in the conversation.

    Tip: For prompts that pull data from multiple systems, check each plugin’s scope of access and the source of the patient or operational information it produces.

  • AvailabilityChatGPT for Healthcare

    Computer-use agents made available for healthcare workflows

    Computer-use agents can perform multi-step tasks in healthcare software, such as reading and clicking on screens, typing, and navigating between systems. The capability is now available for healthcare and administrative workflows through ChatGPT for Healthcare, ChatGPT for Clinicians, and the OpenAI API; access controls, human approval, and audit logs are supported.

    Usage details

    Who can use it: ChatGPT for Healthcare customers, verified individual clinicians in the US, and organizations developing healthcare applications through the API; workspace settings and applicable BAA terms apply.

    How to use it: Current ChatGPT for Healthcare customers should ask their administrator to enable computer use access in Work or Codex; individual clinicians can sign up at chatgpt.com/plans/clinicians; developers should follow the API quickstart and computer use guide.

    Tip: Start the first pilot with a single, reversible workflow that requires human review, and set an approval threshold for every action.

  • APIOpenAI Go SDK

    Go SDK turns Agents session errors into stream events

    Go SDK v3.100.0 reports beta Agents session errors as stream events. In long-running sessions, applications can catch an error during the event stream, not just in the final result.

    Usage details

    Who can use it: openai-go v3.100.0 and later; Agents API beta sessions.

    How to use it: Upgrade the SDK and handle the session error event type separately in your Agents stream loop, notifying the user alongside the progress status.

    Tip: When displaying a stream error, check idempotency and any partial tool effects from the session before retrying.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.100.0 release
  • APIOpenAI Go SDK

    Go SDK adds MCP approvals and Live transport events

    OpenAI Go SDK v3.100.0 adds MCP tool approvals and Live transport events to its typed API surface. This lets applications handle approval flows for MCP calls and Live transfer events through the SDK.

    Usage details

    Who can use it: openai-go v3.100.0 and later; Go applications with access to the relevant beta/Live/MCP features of the API.

    How to use it: In go.mod, upgrade github.com/openai/openai-go/v3 to v3.100.0 or later, then use the compiler to complete the new MCP approval and Live transport event types.

    Tip: Do not set up a default flow that automatically accepts MCP approvals without user or policy approval.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.100.0 release
  • CapabilityOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK added pre-output review and access to accepted text

    The beforeComplete callback lets the application review the proposed output and continue or stop the current turn. turn.acceptedText() exposes accepted text prefixes to the application during generation.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; JavaScript/TypeScript applications using native turn.

    How to use it: Add the beforeComplete callback to the agent or turn configuration, and use acceptedText() during generation for in-app previews or validation.

    Tip: Run output validation before side-effecting operations; treat accepted text as an intermediate signal, not final approval to send.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • CapabilityOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK added LocalStorage profiles and history forks

    Native LocalStorage now supports named profiles, history queries, edits, and forks. The SDK also provides memory, artifact, and UI helpers for sandbox workflows.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; applications using local state and sandboxing.

    How to use it: Create a named profile in the LocalStorage guide, use the history query/edit/fork APIs, and deploy native packages with matching versions together.

    Tip: Back up profile and checkpoint data before upgrading; a matching native Codex revision may be required for restoration.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • CapabilityOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK expanded native voice and Realtime controls

    Native Realtime and voice flows gain new controls for audio input, response timing, history editing, usage information, and SIP or browser call setup across Node.js, browsers, React Native, and Workers.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; supported JavaScript runtimes and relevant voice connections.

    How to use it: Install alpha packages at matching versions, then choose the Node, browser, React Native, Worker, or SIP connection that matches your environment in the Native voice guide.

    Tip: Test history editing and interruption flows in voice sessions with real network latency.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • Plugin / IntegrationOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK added provider options and ChatGPT sign-in routes

    Native provider options now include scoped header/query parameters, WebSocket routes, provider-managed history, buffered Responses JSON, hosted prompts, and retry controls. Native text Sessions also support Sign in with ChatGPT and saved Codex CLI login routes.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; applications using a provider adapter or ChatGPT/Codex authentication flow.

    How to use it: Set provider header/query/WebSocket/history/retry options in Session configuration; provide the appropriate Sign in with ChatGPT or saved Codex CLI login route in your application.

    Tip: Authorize the identity flow separately from model authentication; don't interpret a saved login as automatic trust for all tool permissions.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • AgentOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK adds typed Decisions and DecisionHandoff

    Native Agents can now produce booleans, constrained choices, or supported structured values using agent.decide(), ctx.decide(), and OpenAIDecisions. DecisionHandoff can use these decisions to make structured agent selections for native text handoff.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; Node.js and supported JavaScript/TypeScript environments.

    How to use it: Install the alpha packages at the same version, define a call to agent.decide() or ctx.decide() with explicit input and schema, and use DecisionHandoff for handoff selection.

    Tip: Keep the decision schema small and enum-like to make handoff routing auditable.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • APIOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK added support for application-owned model backends

    Native Agents can work with application-owned model backends through the ResponsesModel, ChatCompletionsModel, and AiSdkModel adapters. Streaming output is also supported within each adapter's limits.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; adapter support may vary by Node.js, browser, React Native, or Workers target.

    How to use it: Install all five SDK packages at the same alpha version, and provide one of the ResponsesModel, ChatCompletionsModel, or AiSdkModel adapters described in the session configuration guide.

    Tip: When choosing a custom backend, test tool calling and streaming events against the official OpenAI backend.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • Plugin / IntegrationChatGPT

    LVMH media content to appear in ChatGPT with sourced summaries and citations

    A strategic partnership between OpenAI and the LVMH Group Media Division was announced. The agreement covers providing summaries and excerpts from LVMH media publications in response to relevant ChatGPT questions, with sources and direct links to original articles; it also includes ChatGPT Enterprise for the teams.

    Usage details

    Who can use it: LVMH Media Division publications and OpenAI's partnership rollout; no specific publication, country, or plan restrictions were announced.

    How to use it: Ask ChatGPT for a sourced summary of a relevant Les Echos, Le Parisien, Challenges, or other LVMH media topic, then open the provided link to the original article.

    Tip: Do not treat the summary as a substitute for the news article; use the direct link to the original article in the ChatGPT response as your primary source.

  • Plugin / IntegrationOpenAI Agents SDK for Python

    Python Agents SDK expands native tool search and MCP controls

    Native tools now offer options for namespaces, concurrency limits, caller constraints, managed approval, and client-side tool search. Native MCP adds discovery filters, cache controls, per-tool approvals, and aliases; read-only workspaces and callback-based execution options have also been expanded.

    Usage details

    Who can use it: v0.24.0-alpha.8 preview; Python applications using native tools and MCP.

    How to use it: Install the alpha, define namespaces and permissions in your tool catalog, configure MCP discovery/approval settings, and defer schemas with ToolSearch when needed.

    Tip: When enabling tool search, keep tool descriptions short and unique; scope approval policies by namespace or tool impact.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • AvailabilityOpenAI Agents SDK for Python

    Python Agents SDK preview adds a Linux ARM64 wheel

    Alpha.8 releases a preview wheel for Linux ARM64. Linux wheels support systems with glibc 2.28 or later.

    Usage details

    Who can use it: v0.24.0-alpha.8; Linux ARM64 and x86_64, glibc 2.28+; other platform requirements depend on the release notes.

    How to use it: Install the Alpha.8 wheel using the pip command appropriate for your platform, and check your glibc version.

    Tip: Before deploying to production, test ARM64 native executor, bubblewrap, and checkpoint compatibility on the target machine.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • CapabilityOpenAI Agents SDK for Python

    Python Agents SDK adds native Realtime audio to Session

    Server WebSocket, browser WebRTC, and existing SIP call connections are supported for adding Realtime audio to native Sessions. Native Realtime APIs are available under agents.realtime.native, while the classic import path is preserved.

    Usage details

    Who can use it: v0.24.0-alpha.8 preview; WebSocket, WebRTC, or SIP connections in non-Node Python applications.

    How to use it: Install the alpha, use the agents.realtime.native API, and configure server WebSocket, browser WebRTC, or an existing SIP call based on the connection type.

    Tip: For voice streams, test tool permissions and interruption/reconnection behavior separately from text streams.

    ✓ Verified by official sourceModels: GPT-LiveSource: Agents SDK v0.24.0-alpha.8
  • APIOpenAI Agents SDK for Python

    Python Agents SDK supports application-owned model clients

    The native tool loop can now work with application-owned model clients, such as async Responses and Chat Completions clients, LiteLLM, AnyLLM, and custom adapters. This lets the application choose its model backend without relying on the SDK's default client.

    Usage details

    Who can use it: v0.24.0-alpha.8 preview; supported Python model clients and adapters.

    How to use it: Install the alpha version, provide your own async client/adapter in NativeProvider or the relevant model configuration, and validate the model loop in a test environment.

    Tip: Before switching to a custom backend, test streaming, tool calling, error formats, and authentication compatibility separately.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • APIOpenAI Python SDK

    Python SDK reports Agents session errors as stream events

    OpenAI Python SDK v3.52.0 exposes beta Agents session errors as stream events. This allows stream-based clients to respond as soon as an error occurs and notify users before a result is returned.

    Usage details

    Who can use it: openai-python v3.52.0 and later; Agents API beta sessions.

    How to use it: Upgrade the openai package and handle the session error event in a separate branch of your Agents stream event handler.

    Tip: When displaying error events to users, also store the session ID and whether the session can be resumed.

    ✓ Verified by official sourceSource: OpenAI Python SDK v3.52.0 release
  • APIOpenAI Ruby SDK

    Ruby SDK turns Agents session errors into stream events

    OpenAI Ruby SDK v0.125.0 reports beta Agents session errors as stream events. Ruby applications consuming streams can handle an error before the final result.

    Usage details

    Who can use it: openai-ruby v0.125.0 and later; Agents API beta sessions.

    How to use it: Update the Gemfile and handle the beta session error event type separately in the Agents stream consumer.

    Tip: Before automatically resuming after an error, check the effects of previous tool calls and whether the session can be safely reopened.

    ✓ Verified by official sourceSource: OpenAI Ruby SDK v0.125.0 release
  • APIOpenAI Go SDK

    Go SDK exposes schedule pause reasons

    Go SDK v3.100.0 adds schedule pause reasons fields that indicate why agent schedules were paused. Applications can now explain a paused schedule with reason details instead of a generic error.

    Usage details

    Who can use it: openai-go v3.100.0 and later; Go applications using the agent scheduling API.

    How to use it: Upgrade the SDK to v3.100.0 and add code to display the pause reason fields in the schedule response to users or use them in a restart policy.

    Tip: Rather than telling users only that a schedule was paused, define a retry, approval, or resolution path for each reason.

    ✓ Verified by official sourceSource: OpenAI Go SDK v3.100.0 release
  • OtherOpenAI Agents SDK for JavaScript

    JavaScript Agents SDK fixed request identity and module/cleanup behavior

    Native HTTP, provider-history, and Responses WebSocket requests now identify the SDK as openai_agents_js and report the installed package version in the User-Agent. ESM/CommonJS identity sharing, Realtime Node entry points, and cleanup/recovery behavior were also fixed.

    Usage details

    Who can use it: v0.22.0-alpha.3 preview; supported Node.js installations and applications using native Realtime.

    How to use it: Update packages to the same alpha version; rerun User-Agent, ESM/CommonJS import identity, and cancellation/recovery tests.

    Tip: If your observability or rate-limit rules depend on the User-Agent, compare logs after upgrading.

    ✓ Verified by official sourceSource: Agents SDK v0.22.0-alpha.3
  • SettingOpenAI Agents SDK for Python

    Python Agents SDK expands request, prompt cache, and model options

    Alpha.8 supports per-request sampling, retry, deadline, and prompt-cache settings. It adds prompt-cache routing for new Sessions with TraceOptions.group_id, prompt_cache_key in the model adapter, and options for Responses JSON/tool-result token limits.

    Usage details

    Who can use it: v0.24.0-alpha.8 preview; Python applications using native model providers and Session.

    How to use it: Add sampling/retry/deadline/prompt cache options to the request or provider configuration; use TraceOptions.group_id for group routing and prompt_cache_key for an adapter override.

    Tip: Do not treat prompt-cache routing as a guarantee of a cache hit; tune it by measuring hit rates and token costs.

    ✓ Verified by official sourceSource: Agents SDK v0.24.0-alpha.8
  • APIOpenAI Python SDK

    Python SDK adds schedule pause reasons

    OpenAI Python SDK v3.52.0 adds reasons for paused agent schedules to the API types. Python applications can present the status and reason in greater detail.

    Usage details

    Who can use it: openai-python v3.52.0 and later; Python applications using the agent scheduling API.

    How to use it: Upgrade the openai package to v3.52.0 or later with pip and handle the pause reason fields in schedule responses.

    Tip: Log pause reasons and send them to the operations team as actionable signals.

    ✓ Verified by official sourceSource: OpenAI Python SDK v3.52.0 release
  • APIOpenAI Ruby SDK

    Ruby SDK adds schedule pause reasons

    OpenAI Ruby SDK v0.125.0 adds reasons for paused agent schedules to the API surface. Ruby applications can handle schedule status more explainably.

    Usage details

    Who can use it: openai-ruby v0.125.0 and later; Ruby applications using the agent scheduling API.

    How to use it: Upgrade the openai dependency in the Gemfile to v0.125.0 or later and handle the pause reason fields in the schedule response.

    Tip: Show users a clear next step and an appropriate restart rule for each pause reason.

    ✓ Verified by official sourceSource: OpenAI Ruby SDK v0.125.0 release

Meta · Meta AI

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

New in this edition · 4

  • FeatureDirect

    Draw, Notes, custom Stickers, Voice Effects and Themes come to Direct

    Direct offers tools to personalize chats, including drawing with Draw, short updates above the inbox with Notes, custom Stickers, Voice Effects and Themes. These tools are rolling out as part of the Android/iOS test of Direct in Türkiye and Thailand.

    Usage details

    Who can use it: Direct test users in Türkiye and Thailand; Android and iOS; features are available with or without an Instagram account.

    How to use it: In a Direct chat, draw with Draw, share a short note in the Notes area above the inbox, create custom stickers with Stickers, use Voice Effects in replies, and customize the chat’s appearance with Themes.

    Tip: Use different Themes and short Notes for people you talk to often to make important chats easier to recognize.

  • AvailabilityDirect

    Instagram’s Direct messaging app is being tested in Türkiye and Thailand

    Meta introduced an independent Direct messaging app offered by Instagram and expanded its test to Türkiye and Thailand. The app is available on Android and iOS with an Instagram account or using an email/phone number without an Instagram account.

    Usage details

    Who can use it: Android and iOS users in Türkiye and Thailand; an existing Instagram account is optional, and no general access beyond the test has been announced.

    How to use it: Install Direct on Android or iOS; sign up with an email/phone number or log in with your existing Instagram username and password. To add someone without an Instagram account, enter their full username or scan their QR code.

    Tip: If you have an Instagram account, log in with it to benefit from having your existing chats synced between the two apps.

  • FeatureMOTHERGUNSHIP: FORGE

    MOTHERGUNSHIP: FORGE adds hand tracking and new game modes

    The “Thank You, Recruit!” update on Meta Quest added full hand tracking, a hand-tracking tutorial, Daily Mode and Ascension Mode with increasing difficulty. The game also moved to Unreal Engine 5.7; the update went live on November 16.

    Usage details

    Who can use it: Live on Meta Quest Store; Touch Plus controllers can still be used; the store listing gives the price as 19,99 ABD doları.

    How to use it: Install or update the game from Meta Quest Store; complete the new tutorial the first time you use hand tracking, then select Daily Mode or Ascension Mode in the game.

    Tip: Quickly assemble weapon parts with hand tracking; switch to Touch Plus controllers for precise aiming or familiar controls.

  • APIWhatsApp Embedded Signup

    More detailed new error codes added for WhatsApp Embedded Signup

    Embedded Signup now returns more descriptive error codes for phone number, verification code and WhatsApp Business app number errors. For example, a number already registered on WhatsApp may now return 3441069 or 3441070 instead of 2494064 in some cases; old codes may also appear during the transition.

    Usage details

    Who can use it: Businesses, solution providers and apps using the Embedded Signup integration; old error codes may continue to work in some cases.

    How to use it: Add codes 3441069 and 3441070 to your error handler, maintain backward compatibility for 2494064, and update message mappings according to the Embedded Signup flow errors documentation.

    Tip: Rather than showing users only a numeric code, log the code alongside the relevant phone/verification step and the request’s trace ID.

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 32

  • ModelAlibaba Cloud Model Studio

    The MindLab Macaron series is now available in Singapore International

    mindlab/macaron-v1.1-venti and mindlab/macaron-v1.1-coding-venti have been added to Model Studio with a 1M context window, while mindlab/macaron-v1-tall offers a faster, lower-cost 256K context window. All three models support OpenAI-compatible and Anthropic-compatible APIs.

    Usage details

    Who can use it: Alibaba Cloud Model Studio Singapore region, International deployment scope; developers with API access.

    How to use it: Select the Singapore region in Model Studio, get the relevant model ID from the catalog, and use it with an OpenAI-compatible or Anthropic-compatible endpoint.

    Tip: Compare the venti versions for long-context Agent and coding tasks, and the tall version when low latency and cost are priorities.

  • APIAlibaba Cloud Model Studio

    Qwen4 API adds automatic and explicit caching options

    The API for Qwen4 now supports automatic caching, a 5-minute or 1-hour TTL, and prompt_cache_breakpoint on specific input blocks. Repeated prefixes can be reused through prompt_cache_options.mode and ttl in the Responses API, or corresponding cache fields in the Chat Completions and Anthropic-compatible surfaces.

    Usage details

    Who can use it: Qwen4 series; Alibaba Cloud Model Studio OpenAI-compatible Chat Completions/Responses and Anthropic-compatible APIs.

    How to use it: Put long, unchanging content at the beginning of the input; set mode and ttl with `prompt_cache_options` in the Responses API, and add `prompt_cache_breakpoint: {"mode":"explicit"}` to the required input_text/input_image/input_file block.

    Tip: Keep the cached prefix at least 1,024 tokens long and send the variable question after the breakpoint; this lets you measure cache hits and cost reductions.

    ✓ Verified by official sourceModels: qwen4-maxSource: Alibaba Cloud Model Studio: Explicit cache (Qwen4)
  • AvailabilityQwen4-Max

    Qwen4-Max comes to QwenWork and QwenStudio alongside QwenCloud

    Qwen4-Max is now available through Token Plan and API on QwenCloud. Alibaba also announced that the model is available on the QwenWork and QwenStudio platforms, extending its deployment across applications beyond the previously known separate Model Studio API access.

    Usage details

    Who can use it: QwenCloud Token Plan/API, QwenWork, and QwenStudio; a platform account and the relevant access plan are required.

    How to use it: Sign in to QwenCloud, QwenWork, or QwenStudio and select Qwen4-Max in the model picker; for API access, use a QwenCloud API key with the qwen4-max model ID.

    Tip: Try the same task in the app and through the API to compare differences in context, tools, and cost.

  • Plugin / IntegrationHiRoute

    Codex Responses Lite can now pass through Gateway with additional tool definitions

    When additional tool declarations are used in Codex Responses Lite requests, HiRoute Gateway can now forward the request to compatible model connections. This allows tool-using Codex workflows to work with more connection types.

    Usage details

    Who can use it: Connections using Codex Responses Lite through HiRoute 0.6.3 Gateway.

    How to use it: Connect Codex to HiRoute Gateway, add the required additional tool declaration fields to the Responses Lite request, and route the request to a compatible model connection.

    Tip: Keep the tool schema small during the initial rollout and verify tool declaration forwarding in Gateway logs.

    ✓ Verified by official sourceSource: HiRoute 0.6.3 sürüm notları
  • APIHiRoute

    HiRoute adds the --apply-file option for routing preview

    hiroute routing preview can now prepare an Apply request to apply Plan lifecycle changes using --apply-file. This makes the CLI step between preview and application more explicit.

    Usage details

    Who can use it: HiRoute 0.6.3 CLI; users managing Plan lifecycles.

    How to use it: Run `hiroute routing preview --apply-file <file>` in the CLI, review the generated Apply request, and apply it with the approved plan.

    Tip: Store the Apply file in version control and review routing differences before applying.

    ✓ Verified by official sourceSource: HiRoute 0.6.3 sürüm notları
  • SettingHiRoute

    HiRoute can restore disabled model routing settings

    When an Agent's model routing is disabled, HiRoute can now restore previously saved settings with confirmation, even if managed settings have changed since then. Unrelated configurations are preserved.

    Usage details

    Who can use it: Agent configurations using HiRoute 0.6.3.

    How to use it: Review and confirm the restore prompt shown when disabling Agent model routing in HiRoute; then check the Agent model settings again.

    Tip: Before confirming the restore, review the differences between the current managed settings and the previous saved version.

    ✓ Verified by official sourceSource: HiRoute 0.6.3 sürüm notları
  • CapabilityHiRoute

    Managed Codex discovers models and recovers subscriptions without a local installation

    Managed Codex sign-in now discovers more available models without requiring a local Codex installation. On Linux and macOS, saved subscriptions are also recovered if a daemon interruption leaves the managed service running.

    Usage details

    Who can use it: HiRoute 0.6.3; daemon interruption recovery on Linux and macOS for users of Managed Codex sign-in.

    How to use it: Upgrade HiRoute to 0.6.3, rerun the Managed Codex sign-in flow, and refresh the model list; after a daemon interruption, sign in again with the same managed subscription.

    Tip: After recovery, verify the account and subscription labels in the model list; don't rely on automatic selection to avoid using the wrong account.

    ✓ Verified by official sourceSource: HiRoute 0.6.3 sürüm notları
  • Plugin / IntegrationQoder

    PHP and Node.js security plugins arrive in Qoder Marketplace

    The Qoder Plugin Marketplace has added php2shell-defense for PHP and nodejs-security-guard for Node.js. The plugins inspect affected runtime paths and prepare a remediation plan and changes after approval.

    Usage details

    Who can use it: Qoder users; desktop/IDE installations with access to Qoder Plugin Marketplace.

    How to use it: In Qoder, open Extensions → Skills, select Qoder Marketplace as the source, search for php2shell-defense or nodejs-security-guard, and click Install; then select the relevant capability with / in a new task.

    Tip: Start with a read-only review prompt; give separate approval for production changes or restarts.

  • AvailabilityHiRoute

    HiRoute better detects Agent installations and subscription credentials on Windows

    Detection of supported Agent installations on Windows is now more reliable. HiRoute can also read subscription authentication updates.

    Usage details

    Who can use it: HiRoute 0.6.3 on Windows and supported Agent installations.

    How to use it: Upgrade HiRoute to 0.6.3 and rerun Agent discovery or the subscription renewal flow; check the detected installation and authentication status.

    Tip: If you have multiple Agent installations, manually verify that the default selection is the correct one.

    ✓ Verified by official sourceSource: HiRoute 0.6.3 sürüm notları
  • SettingQwen Code

    AUTO Mode now requires explicit user authorization for destructive actions

    The AUTO classifier no longer treats an action's relevance to the user's goal as authorization; it looks for an explicit request for destructive, work-deleting, shared-history-rewriting, or bulk actions. It also evaluates questions separately from instructions, along with the scope's blast radius.

    Usage details

    Who can use it: Qwen Code v0.27.4 AUTO Mode; default permission classification for tool calls.

    How to use it: Use AUTO Mode as usual; if you want a destructive or bulk action, state the action and scope explicitly in the prompt rather than specifying only a general goal.

    Tip: Instead of vague phrases like `clean up` or `is it safe`, clearly specify which branch and file are involved and what irreversible operation you want.

  • FeatureQwen Code

    Fixed-interval /loop schedules enabled for Managed Sessions

    Managed Sessions can now define fixed-interval automation with `/loop 5m ...`, list schedules with `/loop list`, and retire them with `/loop clear`. The recovery flow has also been tightened so that only cron tools declared by the session profile are exposed.

    Usage details

    Who can use it: Qwen Code v0.27.4; Managed Sessions using a scheduling profile.

    How to use it: Run `/loop 5m <request>` within a Managed Session; view entries with `/loop list` and remove them with `/loop clear`.

    Tip: Check overlap and catch-up behavior before enabling short intervals in production; choose a meaningful interval to avoid unnecessary calls.

  • SettingQwen Code

    PR status monitoring moved to the operator-only serve.sessionPrWatch setting

    The daemon now manages PR status transitions as the sole poller and configures subscriptions with the operator-only `serve.sessionPrWatch` setting. GitHub Channel’s old status-transition delivery has been removed; `prWatchEvents`, `watch:*`, and related settings remain readable as deprecation aliases for one minor release.

    Usage details

    Who can use it: Qwen Code daemon operator settings; GitHub PR-linked sessions. Old channel keys are retained temporarily for compatibility.

    How to use it: Add `serve.sessionPrWatch.events`, `wakeDormant`, and, if needed, role-keyed `actions` to operator-scoped settings; restart the daemon and check transition events.

    Tip: If you have many linked PRs, account for the effect of `prWatchLimit`; consider setting `prCommentWatch: false` to limit comment traffic.

  • CapabilityQwen Code ACP

    ACP live and read-only Agent Host sessions share tool batches

    ACP live and read-only Agent Host sessions have moved to the shared tool batch path. This means tool execution in these sessions uses the same shared batch behavior as other unified loop surfaces.

    Usage details

    Who can use it: Qwen Code ACP live and read-only Agent Host sessions; nightly 2026-11-15 and later.

    How to use it: Start an ACP Agent Host session with the latest nightly build; select a live or read-only profile and make a tool call to use shared batch execution.

    Tip: Check the profile's capability list first to avoid expecting actual write tools in a read-only profile.

  • CapabilityQwen Code

    ACP prompt, Goal, cron, and notification turns move to the unified loop

    When experimental.unifiedAgentLoop is enabled, ACP prompt and Goal turns, as well as cron and background-notification turns, execute the model request and tool group in a single runAgentLoop call. They can also be moved to the same turn event-bus core with experimental.eventBusAgentLoop.

    Usage details

    Who can use it: ACP sessions using the experimental unifiedAgentLoop; disabled by default.

    How to use it: Enable `experimental.unifiedAgentLoop` in ACP session settings; also enable `experimental.eventBusAgentLoop` to verify the kernel path, then run the ACP prompt/Goal flow.

    Tip: Test the scope of live tool and Agent Host sessions separately; some surfaces in this release retain their own tool step.

  • CapabilityQwen Code

    AgentCore turns run in a single call within unifiedAgentLoop

    When experimental.unifiedAgentLoop is enabled, each AgentCore turn now executes the model request and tool group in a single runAgentLoop call. If experimental.eventBusAgentLoop is also enabled, the turn runs in the event-bus core; default behavior is unchanged.

    Usage details

    Who can use it: AgentCore, subagent, workflow, and teammate flows that enable the experimental setting.

    How to use it: Enable `experimental.unifiedAgentLoop` in Qwen Code settings; also enable `experimental.eventBusAgentLoop` to try the event-bus core.

    Tip: Before production, compare results for the same AgentCore task with the settings off and on; both flags are off by default.

  • CapabilityQwen Code

    AUTO Mode no longer resolves HTTP(S) URLs as incorrect UNC file paths

    In protected write scanning, HTTP(S) URL arguments are no longer sent for filesystem resolution, preventing lengthy UNC/network filesystem probes that could occur on Windows. URLs can still be conservatively classified by AUTO, so protection is not relaxed.

    Usage details

    Who can use it: Qwen Code v0.27.4 AUTO Mode; especially Windows daemon/ACP flows.

    How to use it: Upgrade Qwen Code to v0.27.4; run curl or similar HTTP(S) commands in AUTO Mode without additional settings. Wait for the classifier prompt on protected targets.

    Tip: Account for the additional classifier delay on frequently used safe downloads; do not disable AUTO review for actual write targets derived from URLs.

  • FeatureQwen Code CLI

    CLI preserves feedback input and immediately shows new workflow commands

    Characters in the same stdin batch are no longer lost or sent as the wrong rating in the Ink feedback prompt. Newly saved model-callable workflow commands are also available in the next model request without waiting for the catalog cache.

    Usage details

    Who can use it: Qwen Code v0.27.3-nightly.20261115.6e3b2dcc19 and later CLI/Ink sessions.

    How to use it: Paste the text all at once while the feedback prompt is visible; after saving a workflow command, send a new prompt and check that the command appears immediately in the model catalog.

    Tip: Right after saving a workflow, call the command by itself before starting the first model turn to verify that the catalog has refreshed.

  • SettingQwen Code

    Experimental loop keys can be enforced per surface

    Agent loop keys are now resolved per surface and can be forced on or off with an environment variable. Override formats such as `1`, `0`, and `headless,acp=0` are supported.

    Usage details

    Who can use it: CLI, headless, ACP, and other Qwen Code loop surfaces; experimental configuration.

    How to use it: Set the relevant loop override in the runtime environment; for example, use `1` to enable globally, `0` to disable, and `headless,acp=0` to separate specific surfaces.

    Tip: For a gradual rollout, enable just one surface first; track parity results separately instead of enabling all surfaces at once.

  • CapabilityQwen Code Managed Agent

    Stale public-child recovery now terminates instead of retrying forever

    In Managed Agent, stale public-child recovery is now treated as terminal; the system no longer repeats the same recovery attempt indefinitely. This reduces resource use and wasted wait time for inaccessible child sessions.

    Usage details

    Who can use it: Qwen Code Managed Agent v0.27.3-nightly.20261115.6e3b2dcc19 and later.

    How to use it: Upgrade to the latest nightly or a later version; if recovery fails, open a new child session or restart the task from the parent session.

    Tip: After terminal recovery, create a new child session instead of trying to force the state of the old one.

  • CapabilityQwen Code Hosted Harness

    Hosted turn history is preserved after an interruption

    Hosted turns more reliably continue through the resume flow while retaining their history after an interruption. This reduces the risk that previous turn context is lost or appears incomplete when the user reconnects.

    Usage details

    Who can use it: Qwen Code Hosted Harness nightly 2026-11-15 and later; interrupted and resumed turns.

    How to use it: Interrupt a turn in Hosted Harness, reconnect to the session, and resume it; check that the previous turn history and relevant context are preserved.

    Tip: Before resuming, check the latest tool result and turn status; if they are unclear, don’t blindly restart the same task.

  • FeatureQwen Code Web Shell

    The Web Shell Local files bridge is more visible and takes one step

    The Local files entry now shows connection status with a colored dot and a status/directory tooltip; when connected, Change directory can be done in one step. Actions such as a reauthorization warning, connected session name, New session, copying the daemon/SSH command, and Retry have also been added.

    Usage details

    Who can use it: Qwen Code v0.27.3-nightly.20261115.6e3b2dcc19 and later; users of the Web Shell Local files feature.

    How to use it: In the Web Shell sidebar, open More → Local files; when connected, select Change directory, and use New session if there is no session; if the daemon is unsupported, click Copy startup command.

    Tip: Before giving the Agent access to the local disk, be sure to verify the connected directory and Session information in the tooltip.

  • CapabilityQwen Code ACP

    ACP skips unavailable transcript refreshes during active loading

    ACP skips unavailable transcript refreshes while active loading is in progress, preventing temporarily corrupted or not-yet-ready records from interrupting the session flow. Available records continue to load.

    Usage details

    Who can use it: Qwen Code ACP nightly 2026-11-15 and later; sessions with actively refreshed transcripts.

    How to use it: Open an ACP session with the latest nightly and try sending a new message while the transcript is loading; the session should not drop completely because of unavailable intermediate records.

    Tip: For a persistent transcript error, open a new session and inspect the corrupted record separately instead of forcing the same record.

  • FeatureQwen Code CLI

    Auth dialog exit confirmation is preserved before the window closes

    In the CLI auth dialog, the exit confirmation window no longer disappears during transient state changes. Users can confirm before accidentally exiting the authentication flow.

    Usage details

    Who can use it: Qwen Code CLI nightly 2026-11-15 and later; sign-in flows that use the auth dialog.

    How to use it: Open the auth dialog in the CLI and use the exit shortcut; when the confirmation window appears, choose to continue or cancel.

    Tip: Cancel at the confirmation screen rather than closing the auth dialog before browser-based sign-in is complete.

  • FeatureQwen Code CLI

    The CLI preserves task selection when the task list is reordered

    The visible task selection in the CLI no longer shifts to the wrong item or disappears when tasks are reordered. The selection moves to the same task in its new position.

    Usage details

    Who can use it: Qwen Code CLI nightly 2026-11-15 and later; flows that reorder the task list.

    How to use it: Open the task list, select a task, and perform an action that reorders the list; check that the selection follows the same task.

    Tip: In long workflow lists, verify the task name or ID before starting a bulk operation.

  • FeatureQwen Code CLI

    Multiselect screen correctly shows disabled rows as unselected

    In CLI multiselects, disabled rows are now consistently shown as selected or unselected according to their row state. This prevents inactive options from appearing to be selected by mistake.

    Usage details

    Who can use it: Qwen Code CLI nightly 2026-11-15 and later; multiselect dialogs.

    How to use it: Open the multiselect screen and check the selection indicator for a disabled row; select the active rows and confirm.

    Tip: Instead of copying disabled rows as if they were selectable, first fix the reason selection is disabled.

  • CapabilityQwen Code Web Shell

    Withheld memory content remains read-only

    Web Shell now preserves withheld memory content as read-only without losing it during visibility or loading flows. This prevents the hidden memory segment from being accidentally edited or rewritten.

    Usage details

    Who can use it: Qwen Code Web Shell nightly 2026-11-15 and later; sessions with withheld memory content.

    How to use it: Open the memory or transcript view in Web Shell; withheld content can be viewed, but no editing action is offered.

    Tip: If you need to use hidden memory, copy the content and deliberately create it as a new, visible memory entry.

  • SettingQwen Code

    Hosted unified loop surface added for Hosted Harness

    A new hosted surface is available under experimental.unifiedAgentLoop; Hosted Harness model turns can use the shared runModelRound path. Tool dispatch, turn boundaries, hooks, and recovery flows are preserved, and hosted remains off by default.

    Usage details

    Who can use it: Opt-in Hosted Harness users; the new hosted surface is off by default.

    How to use it: In Qwen Code experimental loop settings, enable unifiedAgentLoop and select `hosted` among the rollout surfaces; then restart the Hosted Harness session.

    Tip: First compare text and recovery scenarios with the old path; the tool batch still runs on the old journal path.

  • SettingQwen Code CLI

    Root-restricted workspace settings are hidden in the CLI

    The CLI no longer presents root-scoped, restricted workspace settings to users as editable options. This prevents attempts to change ineffective or unauthorized settings by mistake.

    Usage details

    Who can use it: Qwen Code CLI nightly 2026-11-15 and later; workspaces with root-restricted workspace settings.

    How to use it: Open the workspace settings; verify that root-restricted settings are not shown, and change settings only within the permitted scope.

    Tip: If a setting is not visible, check its scope and security rule instead of trying to force-edit the file.

  • APIQwen Code SDK

    SDK refreshes replay capability after reconnecting

    When the SDK transport reconnects, replay capability is now refreshed. This prevents the flow for retrieving past events or a transcript after a dropped connection from getting stuck with stale capability information.

    Usage details

    Who can use it: Embedded applications using WebSocket/transport reconnect with the Qwen Code SDK.

    How to use it: Update the SDK to the latest nightly version; deliberately disconnect and reconnect, then retry the replay or transcript refresh call.

    Tip: After reconnecting, call for capability information first, then make the replay request; do not use a stale capability cache on the application side.

  • APIQwen Code TypeScript SDK

    SDK routes Skill changes through the native REST path

    Skill mutation operations in the Qwen Code SDK now use the native REST path. This makes SDK calls such as creating, updating, or deleting a Skill more consistent with the new REST contract.

    Usage details

    Who can use it: Embedded applications using the Qwen Code TypeScript SDK.

    How to use it: Upgrade the SDK to the latest version and run existing Skill mutation calls with the same SDK client; no separate migration flag was specified.

    Tip: After updating, test mutation responses and error cases against the new REST status codes.

  • CapabilityQwen Code CLI

    Buffered output is preserved when a shell process exits

    CLI shell execution no longer loses output left in the buffer just before a process exits. When a command completes, final lines and partial output reach the transcript more reliably.

    Usage details

    Who can use it: Qwen Code CLI nightly 2026-11-15 and later; sessions that run shell commands.

    How to use it: Run normal shell commands in the latest version; check that the final output lines appear in the transcript and tool card even when the process exits quickly.

    Tip: For large outputs, continue to redirect them to a file; the buffer fix does not mean unlimited output storage.

  • CapabilityQwen Code

    Speculation model turns can use the shared agent loop

    When experimental.unifiedAgentLoop is enabled, speculation model turns also move to the shared agent loop; behavior remains unchanged when the setting is off. This expands the scope of experimental loop surfaces in Qwen Code.

    Usage details

    Who can use it: Speculation flows that enable the experimental unifiedAgentLoop.

    How to use it: Enable `experimental.unifiedAgentLoop` in Qwen Code and run a speculation flow; for comparison, turn the setting off and try the same prompt again.

    Tip: Measure speculation results alongside latency and token usage; comparing only the response text is not enough.

    ✓ Verified by official sourceSource: Qwen Code v0.27.4 sürüm notları

Watchlist

Kimi K3.1 Pro

46
Moonshot AI · Kimi

Ranked 11th; still the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. The stable release of Kimi Code 2.6.2 fixed slow starts in long sessions and hangs with empty responses.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; ranked 12th with 44 points. In preview on Google Agent Platform under the name `muse-wonder-1.0`.

Qwen4 Max

43
Alibaba · Qwen

Released on November 16; ranked 13th with 43 points, 9 points above Qwen3.8 Max (34). Available via the API in Model Studio under the name `qwen4-max`, as well as in QwenCloud Token Plan, QwenWork and QwenStudio. 1M context; Singapore pricing is $1,50 / $4,50. The Qwen4 API has supported open caching with a 5-minute or 1-hour TTL since November 16. Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

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

Muse Ember

36
Meta · Meta AI

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

Grok 4.7

35
xAI · Grok

The current model in the Grok app and API. Grok 4.8 is expected to reach some Grok Bot users; its official API ID, pricing and Artificial Analysis score are not yet available. Grok Bot 1.0 launched on November 11.

Sources