YZ·RADARyz.mikail.net
Daily edition · 12 November 2026

Claude Sheets and Whiteboards arrive; Qwen4-Max-Preview enters preview

The TOP 10 is unchanged from yesterday. In Artificial Analysis Intelligence Index v5.0.1, Claude Fable 5.2 leads with 65 points, Gemini 4 Carbon is second with 60, and Claude Opus 5.5 is third with 59; neither the rankings nor the scores changed in the top ten. The watchlist is also unchanged: Muse Wonder 44, Claude Haiku 5.5 and Muse Ember 36, Grok 4.7 and Tencent Hy4 35, Qwen3.8 Max 34. The latest accessible LMArena text ranking is still dated October 27; Gemini 4 Carbon and Gemini 4 Argon lead there. Two new models could affect the rankings in the coming days: Alibaba’s Qwen4-Max-Preview, which opened for preview on November 11, and Grok 4.8, which xAI says it will distribute to a limited group this week. Neither has an Artificial Analysis score yet.

Anthropic delivered the day’s biggest package, focused more on work tools than models. Claude Sheets and Claude Whiteboards opened in beta on November 10. Sheets lets users edit spreadsheets with tabs, formulas, dropdowns, and charts alongside Claude and their team; it imports data from CSV, Excel, or Google Sheets and exports it as XLSX/CSV. Whiteboards bring ideas together on a shared board and turn them into a flow, timeline, document, or presentation. Claude Dashboards, which was in beta, became generally available as Claude Data; it can generate dashboards, data stories, and short narrated videos using a data connector such as Redshift, BigQuery, Snowflake, or Databricks. All three are available on Pro, Max, Ultra, Team, and Enterprise; they’re on by default in Team, while Enterprise admins enable them under Organization settings > Artifacts. Interactive Connectors can now open connected apps in a sidebar or full screen within Claude and edit them; in Team/Enterprise, auto-connect lets an admin authorize a connector once so the team connects automatically. App tools and embedded site capabilities for MCP Apps are now in public preview. Claude Meetings was also announced, but the official page says “coming soon”; only Reuters and TechCrunch report that a Google Meet beta is available for Team and Enterprise. On November 12, a self-hosted version of Claude Managed Agents entered private beta, allowing it to run in your own cloud or on-premises. Another 14 changes were recorded for Claude Desktop, including scheduled task statuses, automatic retries on provider errors, deletion confirmation that can be retained for 30 days, and removal of old tab keys.

At Google, the launches most relevant to developers are Gemini Decision 1 and Decisions API. The model makes yes/no, multiple-choice, and scoring decisions in a single pass over text, image, audio, video, or PDF input; it can answer up to 1,000 questions in one request, and output tokens aren’t charged because it doesn’t generate text. Gemini Live is coming to Apple CarPlay worldwide. A `/swarm` preview opened in Antigravity for Ultra subscribers: a coordinator, parallel specialist agents, and validators work in separate git worktrees. Gemini 4 Carbon was added to Android Studio’s paid tier and the CodeMender preview. On the enterprise side, Agent Gateway entered Limited Preview for managing agents running on Kubernetes clusters outside Google (EKS, AKS, OpenShift); multi-project semantic governance and an agent cost-estimation preview in Cloud Billing also arrived. OpenAI had only Codex CLI 0.167.2: incremental tool notifications are now on by default in Vertex AI sessions. xAI issued no official release notes; Elon Musk wrote that Grok 4.8 will become available to some Grok Bot users this week. Meta released XR All-in-One SDK 209.0. For advertisers, “Measure conversions on customer devices” is now on by default in Meta Pixel for eligible Pixels; Meta’s help page confirms the setting, but the November 12 start date appears only in secondary sources.

Alibaba opened the first model in the Qwen4 family. Qwen4-Max-Preview supports a 1M context, image/video understanding, tool calling, and web search; for now, it’s available only through Token Plan and until November 25 (UTC+8), after which this model ID will be removed. For this model, `max_tokens` limits reasoning and the response together. An Individual tier was added to Token Plan, and the Team plan became about 40% cheaper. Qwen Code 1.0 Preview arrived with bundled local packages that combine Qoder capabilities and don’t require Node.js; HiRoute 0.6.0 added task routing, a Windows desktop app and Linux browser interface, and API key budgets; AgentScope 2.0 and DevOps AI Assistant were updated. In Moonshot, Kimi Code 2.6.0 canary removed the `turn_id` alias from KAP server file-history queries; clients using this name will break. DeepSeek has no new model or pricing, but DeepSeek Harness v0.2.2-rc.2 brought 46 changes: AgentTeam with named and multi-layered SubAgents, code review in Git Worktree, web search via DeepSeek Search API by default, Computer Use that doesn’t move the system mouse, and narrower directory permissions on Windows. At Z.ai, ZCode v3.18.0 added voice input, managed worktrees, Browser design mode, Extension Host, and the free Start Plan provider. Tencent saw 36 changes across LoopForge, BrowserSkill, teamai-cli, AI-Infra-Guard, and Hy-Memory/MemHub. There have been no new changes on Xiaomi’s side since November 10. A total of 159 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 and leads with 65 points. Available on all paid Claude plans (an Enterprise admin must enable it); API ID `claude-fable-5-2`, $8 / $40, with thinking always adaptive. Claude apps added Sheets and Whiteboards (beta) and Claude Data (GA) on November 10.Released: 9 November 2026

  2. 02

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    Generally available since November 10 in Gemini API and AI Studio as `gemini-4-carbon` (paid tier only; $2 / $10 through November 30, $4 / $20 from December 1). Also available in Android Studio’s paid tier; coming to Antigravity for Ultra subscribers with the `/swarm` multi-agent preview.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 via 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 powerful GPT-6 model. The default reasoning effort for this model changed from `low` to `medium` in Agents SDK Python v0.24.0-alpha.7.Released: 3 September 2026

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

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

  6. 06

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

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

  7. 07

    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 admin preview of Dots agents runs through November 18.Released: 26 October 2026

  8. 08

    Gemini 4 Argon

    Google · Gemini=
    50pts

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

  9. 09

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta’s most powerful Muse model; no public API. The Muse app is now also available on Apple Watch and Wear OS.Released: 12 October 2026

  10. 10

    Kimi K3.1 Pro

    Moonshot AI · Kimi=
    46pts

    API $2 / $10; open weights on Hugging Face (2.8T total / 104B active, 1M context); also available on Amazon Bedrock. In Kimi Code 2.6.0 canary, the `turn_id` alias was removed from KAP file-history queries; only `turn_number` is now accepted.Released: 23 October 2026

Key signals

  • The TOP 10 is unchanged: Claude Fable 5.2 leads with 65 points, followed by Gemini 4 Carbon with 60 and Claude Opus 5.5 with 59; the watchlist is also unchanged from yesterday.
  • Anthropic: Claude Sheets and Whiteboards entered beta, and Claude Dashboards became generally available as Claude Data. Interactive Connectors can open connected apps within Claude.
  • Google: Gemini Decision 1 and Decisions API handle up to 1,000 decision questions in a single request, with no output-token charge. Gemini Live is coming to CarPlay, and a `/swarm` preview opened in Antigravity.
  • Alibaba: Qwen4-Max-Preview is available only through Token Plan until November 25; Token Plan added an Individual tier, and the Team plan became cheaper.
  • Breaking changes for developers: `turn_id` was removed from Kimi KAP. DeepSeek Harness rc.2 narrowed Windows directory permissions and switched web search to DeepSeek Search API by default.
  • 159 new developments: DeepSeek 46, Tencent 36, Anthropic 24, Z.ai 17, Alibaba 16, Google 11, Moonshot 5, Meta 2, OpenAI 1, xAI 1. No new changes on Xiaomi’s side.

Worth trying today

01

Move your tracking spreadsheets to Claude Sheets and edit them together

Ask in a chat, “create a monthly tracking sheet from this CSV,” or open Artifacts > Sheets; import the file from CSV, Excel, or Google Sheets. In your first prompt, describe the sheet’s purpose, its columns, and the formulas to calculate, then ask Claude to add dropdowns and a summary chart. Export it as XLSX when you’re done. The feature is on by default in the Team plan; in Enterprise, an admin needs to enable Sheets under Organization settings > Artifacts. For scattered ideas, open Whiteboard from the same place, dictate your notes, and ask it to turn them into a timeline with owners and dates.

Anthropic · Claude
02

Lower the cost of bulk classification with Gemini Decisions API

Get an API key from Google AI Studio, update the SDK (`pip install -U google-genai`), and use the `client.decisions.create(model="gemini-decision-1", input=..., questions=...)` call. Bundle all your yes/no, category, and scoring questions about the same document, image, or recording into one request (up to 1,000 questions); input is charged once, and there’s no output-token charge. This is a good fit for tasks such as tagging support requests, content moderation, or form validation. Route decisions with low confidence scores to human review instead of applying them automatically. The API is currently in v1beta; AI Studio has a free tier for prototyping.

Google · Gemini
03

Try Qwen4-Max-Preview on your own work before the preview ends on November 25

Subscribe to Token Plan (Individual or Team) and create your plan-specific API key, then save it immediately because it won’t be shown again. Set the Base URL in your client to `https://token-plan.ap-southeast-1.maas.aliyuncs.com` and the model to `qwen4-max-preview`; OpenAI Chat Completions, Responses, and Anthropic Messages formats are supported. Since `max_tokens` (or `max_completion_tokens`, but not both) limits reasoning and the response together, leave plenty of headroom even if you expect a short answer. The model ID will be removed after the preview, so don’t tie production workflows to it; use it for comparison testing.

Alibaba · Qwen
04

Use `turn_number` instead of `turn_id` in your Kimi KAP client

Find occurrences of `turn_id` in code that calls the Kimi Code KAP server’s file-history endpoints (`grep -rn "turn_id" .`) and change the parameter to `turn_number`; if no turn number is sent, validation returns an error. Then upgrade the server to 2.6.0 canary and test old and new requests separately. The same release also fixes a reminder received during compact that cancels the operation, a long-path crash on Windows, and loss of context between turns with the Gemini provider. If you use DeepSeek Harness, note before moving to rc.2 that additional directory permissions on Windows no longer cover terminal and PTC access, and that Stagehand’s old `extensionId` setting is invalid.

Moonshot AI · Kimi

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 · 24

  • AvailabilityClaude Managed Agents

    Self-hosted Claude Managed Agents opens as a private beta

    Anthropic has launched a private beta for running the entire Managed Agents service—including the harness, sandbox, credentials, and session history—in a customer’s own cloud or on-premises environment. Model requests can be routed to Anthropic, Bedrock, Google Agent Platform, Microsoft Foundry, or an organization gateway.

    Usage details

    Who can use it: Private beta for selected developers; requires your own cloud or on-premises environment, plus the Managed Agents server, database, and workers.

    How to use it: Request private beta access from Anthropic. Once access is granted, deploy the server, database, and worker components in your environment using the self-hosted Managed Agents setup documentation.

    Tip: Keep session history in your own database and limit model traffic according to your ZDR and organization gateway policies.

  • Plugin / IntegrationClaude auto-connect

    Claude auto-connect brings organization-wide connector setup to beta

    In eligible Team and Enterprise organizations, an administrator authorizes a connector once, and team members connect automatically the first time they use it instead of signing in individually. IdP-managed authorization is generally available; Claude-issued authorization is in beta.

    Usage details

    Who can use it: Eligible Team and Enterprise plans; role- and scope-based management, with IdP-managed or Claude-issued identity sources depending on provider support.

    How to use it: Go to Organization settings > Connectors > connector > Configuration > Managed authorization > Set up, choose the identity source, roles, and scopes, then click Save & turn on.

    Tip: Pilot the setup by enabling the connector for just one custom role before rolling it out to the entire organization.

  • AvailabilityClaude Data

    Claude Dashboards becomes generally available as Claude Data

    Claude Dashboards, previously in beta, has been renamed and is now generally available as Claude Data. Alongside dashboards, it can produce data stories, visual analyses, and short narrated videos.

    Usage details

    Who can use it: Pro, Max, Ultra, Team, and Enterprise; requires a data warehouse or a compatible application connector. It is enabled in Team; the default rollout for Enterprise is 2026-11-19.

    How to use it: Open Artifacts > Data or request a data output in a chat. First, add Amazon Redshift, BigQuery, ClickHouse, Databricks, Snowflake, or another data connector.

    Tip: Open the SQL query Claude generated for each chart and check that the number comes from the correct tables and filters.

  • FeatureClaude Meetings

    Claude Meetings announced as a live-work feature for video calls

    Anthropic’s official post announces Claude Meetings, which can join a meeting and produce notes, plans, and documents, but describes the feature as coming soon. Reuters and TechCrunch reported initial beta access for Team and Enterprise customers on Google Meet; Anthropic has not confirmed this access on its official help page.

    Usage details

    Who can use it: Coming soon, according to the official source; Google Meet beta for Team and Enterprise, according to Reuters and TechCrunch.

    How to use it: Add Claude as a participant when the process for adding it to a Google Meet meeting is available; official sources currently provide no activation steps.

    Tip: For meetings with external guests, check administrator approval and participant notification requirements in advance.

  • FeatureClaude Sheets

    Claude Sheets enters beta with collaboratively editable spreadsheets

    Claude Sheets lets users edit spreadsheets with Claude and their team, including tabs, formulas, formatting, dropdown lists, and charts.

    Usage details

    Who can use it: Pro, Max, Ultra, Team, and Enterprise; full use on web and Desktop, with basic read-only access on mobile. On by default in Team; in Enterprise, enable it under Organization settings > Artifacts.

    How to use it: Request a spreadsheet in a chat or open Artifacts > Sheets. You can import data from CSV, Excel, or Google Sheets and export the result as XLSX or CSV.

    Tip: In your first prompt, specify the sheet’s purpose, the fields to track, and the existing figures. Export to XLSX if you need complex Excel features.

  • FeatureClaude Whiteboards

    Claude Whiteboards enters beta for collaborative team brainstorming

    Claude Whiteboards lets teams organize ideas with Claude on the same board, create flows and timelines, and turn the result into a document or presentation.

    Usage details

    Who can use it: Pro, Max, Ultra, Team, and Enterprise; web, Claude Desktop, and Claude Mobile. On by default in Team; in Enterprise, enable it under Organization settings > Artifacts.

    How to use it: Request a board in a chat or open Artifacts > Whiteboard. Then edit the board, share it with Share, or transfer it to a supported tool with Send to.

    Tip: Start by dictating unstructured ideas, then ask Claude to group them into themes and turn them into a timeline with owners and dates.

  • Plugin / IntegrationInteractive Connectors

    Interactive Connectors can open and edit the full app within Claude

    Supported connectors can now open beside the conversation in a side panel or full-screen; the user and Claude can work on a page, row, or file in the same app. Connectors can also be opened without navigating to them and pinned to the sidebar.

    Usage details

    Who can use it: On all plans; side panel, full-screen, and pin features are available on the web and in Claude Desktop with the new Claude experience. Interactive cards are available on mobile, but full site view is not yet available.

    How to use it: Connect a connector with the Interactive badge in Connectors Directory; then open it or pin it from the + menu, and use Claude's approval cards for actions.

    Tip: For connectors that can make changes, start with a narrowly scoped request and keep Claude's action approvals enabled.

  • CapabilityClaude Desktop ve Claude Cowork

    Additional safety checks added for Auto mode and Cowork shell operations

    When an admin policy disables Auto mode, active sessions now exit it and show the reason. Code dev-server launches in organization configuration require approval, and Cowork Auto mode checks before each shell command.

    Usage details

    Who can use it: Managed Claude Desktop/Cowork/Code deployments; behavior ships with v5.1231.0.

    How to use it: Set Auto mode and tool policies in managed settings; users can review blocked or approval-pending steps from the session card.

    Tip: Test the security approvals required for dev-server and shell commands in automated tasks ahead of time.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • CapabilityClaude Data

    Claude Data adds beta capabilities for certified queries and live sharing

    Data teams can certify queries so everyone uses the same metrics. Live sharing can show approved query results to team members who don’t have access to the connector, keeping them up to date.

    Usage details

    Who can use it: In beta on Claude Data; the relevant features are on by default in Team, while certified queries and live sharing are off by default in Enterprise.

    How to use it: In the query panel, follow Certify query > add instructions > Test > Publish. To enable live sharing, open the Sources menu in the artifact and select Your access for the relevant connector.

    Tip: Before enabling live sharing, check the results shown on the dashboard and the scope of your connector access.

  • CapabilityClaude Data

    Claude Data adds support for design systems, caching, and version rollback

    Claude Data can now use design systems, cache dashboard refreshes and manage them with per-dashboard cadence, and restore previous versions using tracked changes.

    Usage details

    Who can use it: Pro, Max, Ultra, Team, and Enterprise plans that include Claude Data; these capabilities are generally available, except for certified queries and live sharing.

    How to use it: Open a Data artifact; use the design system options, Refresh data settings, or version history panel, and click Restore when needed.

    Tip: Set refresh intervals for frequently opened dashboards based on how quickly the data changes to reduce unnecessary query costs.

  • CapabilityClaude Cowork

    Instruction, comment, approval, and model-switch behavior fixed in Cowork tasks

    When Cowork changes a session's CLAUDE.md file, it no longer changes global instructions. File comments, copied message characters, read-only tasks on unsupported computers, and the Automatically approve transition after model selection now work more reliably.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; Cowork tasks.

    How to use it: Use Settings for global instructions; leave task-specific file comments on the file, and check the Automatically approve toggle after selecting a model.

    Tip: When reopening imported sessions, check the Go back or Continue choice on the first message card.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • RetiredClaude Desktop

    Desktop Home setting controls the unified experience; legacy tab toggles are being removed

    In managed setups where only Code is enabled, unset desktopHome can remain Code-only after November 10. The Chat, Cowork, and Advanced file analysis toggles in the Setup window have been replaced by the Unified Claude setting; legacy toggles will be ignored on December 17.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; managed setups. Warnings for legacy toggles begin on 2026-11-17, and they will be ignored starting 2026-12-17.

    How to use it: Explicitly set desktopHome to off or standard in the configuration; use the new setting instead of coworkTabEnabled, chatTabEnabled, and chatAdvancedFileAnalysisEnabled.

    Tip: Remove legacy toggles and explicitly deploy desktopHome before rolling out the upgrade across your organization.

  • SettingClaude Desktop ve Claude Code

    Desktop MCP servers now follow Code management policies

    MCP servers added to the Desktop app now comply with Claude Code's managed allow list, deny list, and plugins-only policy. In the same release, when an admin policy disables Auto mode, active sessions exit it; freezes in the Code tab across many sessions and some SSH or worktree reconnection issues are also fixed.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; Code sessions using managed Claude Code settings.

    How to use it: Have the admin update MCP permissions in managed settings; users can see which servers are retained in the MCP list in a Code session. Auto mode changes take effect at the next configuration check.

    Tip: Before deploying the policy, verify the MCP allow/deny list and Auto mode behavior on a test device.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • APIMCP Apps

    App tools and embedded site capabilities for MCP Apps enter public preview

    In the MCP Apps extension, Claude supports app tools, entry points, embedded websites, messages, and downloads, in addition to side panel and full-screen views. Developers can add interactive interfaces to their own MCP servers.

    Usage details

    Who can use it: Public preview for MCP server developers; app tools, entry points, and embedded websites require a remote server and a verified directory listing.

    How to use it: Add an interface to your server with the MCP Apps SDK; connect the server in Claude Desktop via Settings > Developer > Edit Config, then restart the app.

    Tip: Pin server and SDK versions; test the required domain and verified listing conditions for embedded sites before deployment.

  • SettingClaude Desktop

    MCP platformOverrides and automatic import behavior expanded

    Managed MCP servers can now run command only on specified operating systems using platformOverrides. Claude's import flow preserves Cowork project descriptions, project and session folders; the sessionStatusLine program can also receive cost information.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; MDM or bootstrap configurations.

    How to use it: Define command under win32, darwin, or linux in the managedMcpServers entry; configure automatic3pImport for imports, and inferenceModelPricingEnabled with sessionStatusLine for cost information.

    Tip: Verify platform-specific command fallbacks on a test device first; older Desktop versions may drop the entry.

    ✓ Verified by official sourceSource: Claude Desktop configuration changelog
  • APIClaude Desktop

    Model identity, scheduled task retries, and limit accounting fixed

    Model IDs in inferenceModels now match only the version specified; sessions saved with an unavailable model can fall back to the default model. Scheduled tasks are retried when the provider returns 500 or 529, and inferenceMaxTokensPerWindow accounting now includes subagents and compaction tokens.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; managed model lists and scheduled tasks.

    How to use it: Use exact version IDs in the model list; move tasks to a current default model and monitor limit changes using provider usage reports.

    Tip: Using an exact model ID instead of an alias prevents a silent model change after an upgrade.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • Plugin / IntegrationClaude Desktop on Vertex AI

    Local Google Cloud credentials required for Vertex AI Sign in with Google flow

    Hooks, MCP servers, and Code tab commands using Sign in with Google on Vertex AI now use the device user's Google Cloud credentials.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; Code and managed tool sessions using Google Cloud or Vertex AI.

    How to use it: Run gcloud auth application-default login on the device, then restart the Claude Desktop or Code session.

    Tip: On corporate devices, verify that the ADC identity has the correct Google Cloud project and permissions.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • SettingClaude Desktop

    Windows data path and computer use security behavior changed

    New Claude Desktop installations on Windows and eligible existing data now use the %LOCALAPPDATA%\Claude-Data path. Computer use now requests Use system shortcuts permission before pressing Windows shortcuts; it also no longer opens file types that install software or apply system settings when double-clicked.

    Usage details

    Who can use it: Windows 10 21H2 and later and Windows Server 2022; Microsoft Store versions and third-party distributions are excluded from the data path change. The computer use permission check is in the Windows client.

    How to use it: Upgrade Claude Desktop to 5.1231.0 or later; grant Use system shortcuts permission when prompted for computer use. Open blocked files with a trusted management tool instead of asking Claude to run them directly.

    Tip: Update your corporate backup and exclusion rules for the new Claude-Data directory.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • SettingClaude Desktop

    Setting added to display organization announcements on managed devices

    Admins can now distribute an organizationAnnouncement message that appears as a dismissible card or modal on the startup screen. The announcement is shown to users again if its ID or text changes.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; devices managed with MDM or bootstrap.

    How to use it: Add the title, body, placement, and id fields under appearance.announcement in the configuration; update all devices first so older devices do not reject the policy because of an unknown key.

    Tip: Change the announcement ID to a version or campaign identifier to make important messages appear again.

    ✓ Verified by official sourceSource: Claude Desktop configuration changelog
  • SettingClaude Cowork

    File deletion approval for scheduled Cowork tasks can be retained for 30 days

    Approval given before a scheduled Cowork task deletes files in its own folder can now be retained for 30 days. Admins can change the duration with mcpScheduledTaskApprovalLifetimeDays or disable the option by setting it to 0.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; Cowork tasks using managed 3P configuration.

    How to use it: Select Allow for 30 days on the task's first deletion request; for the admin setting, deploy mcpScheduledTaskApprovalLifetimeDays as a number of days.

    Tip: For large folders, use a shorter duration or one-time approval instead of 30-day approval.

  • FeatureClaude Desktop

    Scheduled task cards now show detailed run statuses

    In Claude Desktop, Cowork, and Code, scheduled tasks now clearly indicate whether the last run failed, will be retried, is waiting for input, or did not run. Notifications also distinguish a skipped step from an Auto mode safety block.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; Cowork and Code scheduled tasks running on the computer.

    How to use it: Open a task card on the Scheduled tasks screen; view the new status labels and last-run details directly on the card.

    Tip: For Needs input, open the task and grant the pending approval; for recurring errors, check provider or connector access.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • CapabilityClaude Desktop

    Scheduled tasks now retry after provider errors

    When a provider or gateway returns a server error such as 500 or 529, scheduled tasks retry before Claude responds; network or token refresh issues report task status more accurately.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; local Cowork and Code scheduled tasks.

    How to use it: No changes are needed in task settings; monitor the initial error and final result on the Scheduled tasks card, and fix gateway token refresh.

    Tip: For long tasks, wait for the retry status so you don't mistake the initial error notification for a final failure.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • CapabilityClaude Desktop

    Fixed artifact, browser, sign-in, and startup reliability in Desktop

    Fixed files going missing when files with the same name were written to Downloads simultaneously, artifacts reverting to an older version after Refresh, Linux crashes after hiding the browser, inability to access the browser page in some sessions, and Windows installation and Run on startup issues. Email links after failed sign-in and refreshes of registered sessions receiving proxy or portal responses were also fixed.

    Usage details

    Who can use it: Claude Desktop v5.1231.0 and later; relevant client flows on Windows, macOS, and Linux.

    How to use it: Update Claude Desktop to 5.1231.0 or later; no additional settings are required. If the issue persists after updating, get a diagnostic report from Help > Troubleshooting.

    Tip: Save open artifact edits before updating, and add sign-in refresh URLs to your corporate proxy's allowlist.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10
  • Plugin / IntegrationClaude Code

    Pull request checks fixed behind GitHub Enterprise proxies

    Fixed HTTP 401 failures in pull request status, checks, and auto-fix operations on GitHub Enterprise Server installations where an authenticated proxy can be used through GitHub CLI.

    Usage details

    Who can use it: Code sessions in Claude Desktop; organizations using GitHub Enterprise Server and an authenticated proxy.

    How to use it: Upgrade Claude Desktop to 5.1231.0 or later and keep GitHub CLI proxy authentication working on the device.

    Tip: First, verify status and checks calls on a test PR.

    ✓ Verified by official sourceSource: Claude Desktop changelog 2026-11-10

Google · Gemini

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

New in this edition · 11

  • FeatureGemini Enterprise Agent Platform

    Agent Gateway can manage non-Google Kubernetes agents

    Agent Gateway for cross-cloud launched as a Limited Preview, connecting agents running on CNCF-compliant Kubernetes clusters outside Google Cloud to centralized Agent Platform governance. Agent Registry, identity, LLM/MCP/A2A traffic, Model Armor policies, and observability can all be applied from a single control plane.

    Usage details

    Who can use it: Limited Preview for CNCF-compliant clusters outside GKE, such as Amazon EKS, AKS, Red Hat OpenShift, and bare metal. Gateway is free during the preview; Model Armor, Cloud Logging, and Cloud Trace are billed at standard rates.

    How to use it: Fill out the preview access form; then install the cross-cloud Agent Gateway components in your cluster, register agents with Agent Registry, and route proxy traffic through your existing service mesh.

    Tip: Before production, verify identity federation, outbound access to Google APIs, and the preview's known fail-open behavior in a separate test cluster.

  • APIGemini API

    Gemini Decision 1 and Decisions API released

    Google released Gemini Decision 1, a model that makes binary, multiple-choice, and scoring decisions on text, image, audio, video, and PDF inputs in a single pass, along with the Decisions API. A single request can include up to 1,000 independent questions; output tokens are not billed because the model does not generate output text.

    Usage details

    Who can use it: Available in Gemini API and Google AI Studio; a free usage tier is available in AI Studio for prototyping. Accessible through API v1beta.

    How to use it: Get an API key from Google AI Studio; update the SDK with `pip install -U google-genai` or `npm install @google/genai@latest`, then use the `client.decisions.create(model="gemini-decision-1", input=..., questions=...)` call.

    Tip: Combine independent questions about the same input in one request to pay for the input only once; route low confidence values for human review.

  • Plugin / IntegrationGemini

    Gemini Live is coming to Apple CarPlay globally

    The Gemini app is becoming available on Apple CarPlay; iPhone users can have hands-free, two-way conversations with Gemini Live from their car's display. Users can continue previous chats, and conversations sync with Gemini history on their iPhone.

    Usage details

    Who can use it: Rolling out globally to all Gemini users on iOS in supported Gemini Live languages; requires a CarPlay-compatible vehicle and the latest Gemini iPhone app.

    How to use it: Update the Gemini app on your iPhone, grant microphone permission, and use Gemini Live on your phone at least once. After connecting to CarPlay, tap the Gemini icon on the car's display to start a new or recent chat.

    Tip: Use short voice requests while driving; enable Daily Brief beforehand for email and calendar summaries.

  • OtherCodeMender

    CodeMender v0.21.0 improves sandboxing and error resilience

    CodeMender v0.21.0 runs build commands inside the sandbox when sandboxing is enabled, and makes host toolchains and package caches read-only and session-isolated. It adds a 30-minute timeout for silent sessions, preserves evidence output on errors, and automatically retries when sandboxing fails.

    Usage details

    Who can use it: For CodeMender v0.21.0 users in Gemini Enterprise Agent Platform; applies to the `cm build`, `cm fix`, and `cm verify` workflows.

    How to use it: Update CodeMender to v0.21.0; enable sandboxing, run `cm build`, `cm fix`, or `cm verify`, and review the verification output.

    Tip: In large repositories, start by running `cm verify` in sandbox mode; patch refreshes and the 30-minute silent-session limit make debugging easier.

  • AvailabilityCodeMender

    CodeMender supports Gemini 4 Carbon in preview

    CodeMender added Gemini 4 Carbon to its available models as a Preview. Gemini 3.8 Flash remains the default model; Carbon can be explicitly selected for difficult code review and repair tasks.

    Usage details

    Who can use it: CodeMender Preview in Gemini Enterprise Agent Platform; model support is available in the `cm find`, `cm verify`, and `cm fix` commands.

    How to use it: Run one of the commands with the `--model gemini-4-carbon` option: `cm find --model gemini-4-carbon`, `cm verify --model gemini-4-carbon`, or `cm fix --model gemini-4-carbon`.

    Tip: Reserve Carbon for long and complex verification tasks; the default Gemini 3.8 Flash may be faster and more economical for routine work.

    ✓ Verified by official sourceModels: Gemini 4 CarbonSource: Gemini Enterprise Agent Platform — Release notes, 11 Kasım 2026
  • CapabilityGemini Intelligence

    Pixel Solver brings multi-step calculations to Pixel 11

    According to Mashable's November 11 report, Pixel 11 users can use Gemini-powered Pixel Solver to turn written prompts into calculation templates and solve multi-step problems. I couldn't confirm this detail on Google's official Pixel Drop page dated November 10, which I opened.

    Usage details

    Who can use it: Pixel 11 only, according to the independent report; associated with the November 2026 Pixel Drop rollout.

    How to use it: Install the November Pixel Drop update on Pixel 11 and open Pixel Solver to describe your calculation problem in natural language; check the resulting template and answer.

    Tip: Since this feature has not been confirmed by an official Google source, independently recalculate results for critical calculations and wait for the official release notes.

  • FeatureGemini Enterprise Agent Platform

    Multi-project semantic governance arrives in Agent Platform

    Semantic governance policies can now be centrally enforced in a hub-and-spoke, multi-project architecture. A Resource Manager boundary with the `AGENT_MANAGEMENT` capability automatically discovers agents in member projects in the central Agent Registry and manages them through a single Agent Gateway and policy engine.

    Usage details

    Who can use it: For agents deployed across multiple Google Cloud projects; requires a central management project and a Resource Manager management boundary between the member projects.

    How to use it: Create a central management boundary and enable the `AGENT_MANAGEMENT` capability; set up Agent Registry, Agent Gateway, and the semantic governance policy engine in the central project, then connect the member projects.

    Tip: Before expanding policies, test identity, traffic, and violation logs in one member project using narrowly scoped rules and observation mode.

  • AgentGoogle Antigravity

    Multi-agent orchestration preview with /swarm begins in Antigravity

    Alongside Gemini 4 Carbon, Google Antigravity has opened a `/swarm` preview that runs teams comprising a coordinator, parallel specialist agents, and adversarial validators in isolated git worktrees. The new Board, Members, and Graph views let users track team discussions, tool activity, and result verification flows.

    Usage details

    Who can use it: Starting today, Google AI Ultra subscribers get Gemini 4 Carbon access in Antigravity App, CLI, and IDE extensions; the `/swarm` preview is rolling out in App and CLI. Separate access conditions apply to enterprise and scientific use.

    How to use it: Type `/swarm` in the Antigravity App or CLI composer; in CLI, open the live board with `/swarm board` or `Alt+S`. You can pin the coordinator or validator to Gemini 4 Carbon from the member list.

    Tip: Assign Carbon to coordinator and validator roles, and faster models to file-editing workers; multi-agent work consumes quota much faster than standard chat.

    ✓ Verified by official sourceModels: Gemini 4 Carbon, Gemini 4 ArgonSource: Google Antigravity — Gemini 4 Carbon and /swarm
  • AgentCloud Billing

    Cost estimator agent for Gemini agents in Cloud Billing enters preview

    Cloud Billing has begun offering Cost estimator agent as a Preview feature. It provides high-level cost estimates based on reference architectures for AI agents built with Gemini Enterprise Agent Platform. The estimates help with pre-launch budget, spending limit, and approval planning.

    Usage details

    Who can use it: Preview in the Billing section of Google Cloud Console; available for Gemini Enterprise Agent Platform reference architectures.

    How to use it: Open the Billing section in Google Cloud Console, select Cost estimator agent, and generate an estimate by specifying your planned Gemini Enterprise Agent Platform agent architecture.

    Tip: Use the estimate alongside your pre-launch budget and alert thresholds. Treat the results as high-level reference estimates and update them with actual usage measurements.

  • AvailabilityAndroid Studio

    Gemini 4 Carbon becomes available to paid-tier users in Android Studio

    Gemini 4 Carbon is now available to developers in Android Studio through the paid tier of the Gemini API. In addition to being generally available in AI Studio, the model can now be selected directly in the Android development environment.

    Usage details

    Who can use it: Available in Google AI Studio and Android Studio for developers using the paid tier; the model is offered at promotional pricing through 30 November 2026.

    How to use it: Connect your paid-tier API key to Android Studio and select `gemini-4-carbon` in the model picker. For model details, refer to the Gemini 4 Carbon Developer Guide.

    Tip: In Android Studio, reserve Carbon for long planning and complex debugging tasks, and switch to a faster model for short code-completion tasks.

    ✓ Verified by official sourceModels: Gemini 4 CarbonSource: Google Blog — Gemini 4 Carbon is now available for developers
  • OtherGemini CLI

    Gemini CLI 0.69.0-preview.0 fixes session and terminal bugs

    Gemini CLI v0.69.0-preview.0 improves Windows PowerShell policy detection, `@file` paths with spaces, piped stdin, ACP retries, and lengthy network-related waits, separately from the `/stats heatmap`, theme, model, and timeout changes in the known records. It also fixes model-turn 400 errors during resume, headless abort behavior, repeated interactive sessions, and some multipart tool output issues.

    Usage details

    Who can use it: Gemini CLI pre-release v0.69.0-preview.0, dated 10 November 2026; for CLI users on the preview channel.

    How to use it: Select the `v0.69.0-preview.0` tag on the GitHub Releases page and update your existing preview installation to this version; test headless, ACP, shell, and resume workflows in particular.

    Tip: Try the preview release in an environment separate from production, and back up your session history and CLI configuration files before major upgrades.

OpenAI · ChatGPT

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

New in this edition · 1

  • AvailabilityCodex CLI

    Incremental tool notifications for Vertex AI are now the default in Codex CLI 0.167.2

    With Codex CLI 0.167.2, tool updates are now sent incrementally by default in Google Cloud Vertex AI sessions. Unlike the previously recorded general GCP tool-notification expansion, this is a change to the default behavior; an explicit option is available to revert to the old behavior.

    Usage details

    Who can use it: Google Cloud Vertex AI users running Codex CLI 0.167.2; other providers are not affected by this specific default change.

    How to use it: Update with `npm install -g @openai/codex@0.167.2`. To restore the old behavior, add `incremental_tools_legacy_gcp = false` to the `[features]` section in `config.toml`.

    Tip: Keep the default if you want to see live tool progress; temporarily disable this option if your older integration does not process incremental events.

Meta · Meta AI

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

New in this edition · 2

  • SettingMeta Pixel / Events Manager

    Conversion measurement on customer devices enabled by default in Meta Pixel

    Meta’s official Help Center confirms that the `Measure conversions on customer devices` setting is enabled by default for eligible Meta Pixel installations, and that measurement takes place on the customer’s device alongside the existing Pixel and Conversions API setup. The Help page does not specify when it was enabled; Bolstered Media and Jon Loomer Digital report that this behavior took effect on November 12, 2026, so confirmation of the date/rollout is partial.

    Usage details

    Who can use it: Eligible Meta Pixel datasets; Pixel must be installed directly on the webpage, not in an iframe or as an image-only Pixel. Measurement is not enabled for people in the EU and UK.

    How to use it: Go to Meta Events Manager > Data sources > Pixel name and ID > Settings > Meta Pixel settings. Turn the `Measure conversions on customer devices` toggle on or off; it is on by default for eligible Pixels.

    Tip: Compare purchases reported by Meta before November 12 with store/CRM orders to establish a baseline; distinguish subsequent changes in ROAS or CPA from increases in actual orders, and do not remove Conversions API.

  • APIMeta XR All-in-One SDK (UPM)

    Meta XR All-in-One SDK (UPM) 209.0 released; Unity 6000.0.66f2 required

    Meta released version 209.0 of Meta XR All-in-One SDK (UPM) on November 10. The package bundles current versions of Meta XR SDKs as dependencies and requires Unity 6000.0.66f2 or later.

    Usage details

    Who can use it: Developers building Meta Horizon/Quest VR applications with Unity; package 209.0 is available for Unity 6000.0.66f2 or later.

    How to use it: In Unity, go to Window > Package Manager > + > Add package by name…; add the `com.meta.xr.sdk.all` package and select version 209.0. If you need samples, install the samples for the dependent SDKs pulled in by the package separately.

    Tip: Test the upgrade on a separate branch; the package updates dependencies such as Meta XR Core, Audio, Haptics, Interaction, Platform, Voice, and MR Utility Kit together.

Moonshot AI · Kimi

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

New in this edition · 5

  • RetiredKimi Code KAP

    The `turn_id` alias was removed from file-history queries

    The KAP server removed the `turn_id` alias from file-history queries. Clients now need to use `turn_number` as the query parameter.

    Usage details

    Who can use it: Kimi Code KAP file-history API clients; 2.6.0-canary.20261111095258-5518d54 and later versions.

    How to use it: Remove the `turn_id` parameter from the client code and send `turn_number` instead; then upgrade the Kimi Code server to the relevant canary version.

    Tip: Before deploying to production, test the old and new query formats separately to catch backward-compatibility issues.

  • CapabilityKimi Code

    A reminder received during compact no longer cancels the operation

    In Kimi Code 2.6.0-canary, a reminder received while the Agent is compacting no longer cancels the compact operation. This preserves ongoing compaction when a reminder message arrives during a long session.

    Usage details

    Who can use it: Kimi Code 2.6.0-canary.20261111095258-5518d54 and later versions.

    How to use it: Upgrade Kimi Code to the specified canary version; send a reminder while the Agent is compacting and observe that the operation completes without being canceled.

    Tip: During long sessions, you can use reminders to send context corrections while compact is in progress; still, minimize unnecessary reminders.

  • CapabilityKimi Code

    Gemini context is preserved across turns

    A problem that caused context to be lost across turns in sessions using the Gemini provider in Kimi Code has been fixed. Multi-turn flows now continue while preserving previous context.

    Usage details

    Who can use it: Kimi Code multi-turn sessions using the Gemini provider; 2.6.0-canary.20261111095258-5518d54 and later.

    How to use it: Upgrade Kimi Code to the relevant canary version and start a session with the Gemini provider; continue subsequent requests in the same session to verify that previous context is preserved.

    Tip: For long tasks with Gemini, continue the same session instead of starting a new one; this carries over previous decisions and file context.

    ✓ Verified by official sourceModels: GeminiSource: Kimi Code 2.6.0-canary 11 Kasım 2026 sürüm notları
  • CapabilityKimi Code

    Windows batch scripts are now launched with `cmd.exe`

    Kimi Code now runs Windows batch script commands through `cmd.exe`. This fixes an incompatibility when the Agent launches `.bat` and `.cmd` scripts.

    Usage details

    Who can use it: Kimi Code users on Windows who use `.bat` or `.cmd`; 2.6.0-canary.20261111052659-f26481d and later versions.

    How to use it: Upgrade Kimi Code to the relevant canary version and have the Agent run a `.bat` or `.cmd` command in a Windows workspace; the script will be launched through `cmd.exe`.

    Tip: Keep the command in an explicit `.cmd` file to verify the environment variables used by the script and its exit code.

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 1

  • AvailabilityGrok Bot

    Limited rollout of Grok 4.8 planned for Grok Bot users

    Elon Musk announced that Grok 4.8, whose pre-training was completed on November 10 and which is now in the reinforcement learning phase, will have a limited soft rollout for Grok Bot users later in the week. This is not a general release or an API launch, but a plan for controlled user access; the model is not yet publicly available.

    Usage details

    Who can use it: A limited group of Grok Bot users; the post did not specify an exact day, country, plan, or selection criteria. At the time of checking, Grok 4.7 was the latest model in the official developer model catalog; API access for Grok 4.8 was not announced.

    How to use it: Sign in to Grok Bot and wait for access to be enabled for your account; the official post did not provide a manual activation step. For API use, do not try to call Grok 4.8 with an API request at this stage, as no model ID has been published.

    Tip: Continue using grok-4.7 for production integrations; if Grok 4.8 becomes available on your account, start with limited testing and avoid creating a fixed dependency until its official model ID, pricing, and documentation are published.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 36

  • FeatureAI-Infra-Guard

    AI-Infra-Guard releases local generative AI security self-auditing

    The new `aig-compliance` package runs refusal checks and infrastructure scans with target, judge, and attacker models on a local web page, then generates a printable report. API keys are not saved to the local database; DiagSafe, DiagBase, and DiagAttack evaluations are supported.

    Usage details

    Who can use it: Local operators using Python 3.12 or later; model endpoints can be OpenAI-compatible.

    How to use it: Run `pip install aig-compliance`, then `aig-compliance`; enter the target, judge, optional attacker model, and infrastructure addresses at `http://127.0.0.1:8765`.

    Tip: Do not put real keys in URLs or reports; the tool can use the `TARGET_*`, `JUDGE_*`, and `ATTACKER_*` environment variables to fill in the fields.

  • OtherAI-Infra-Guard

    AI-Infra-Guard v4.6.10 adds 135 AI product security advisories from NVD

    The v4.6.10 release notes list a total of 135 AI product CVE advisories from NVD and setup for local security self-auditing. The release also fixes browsers handling asynchronous local tool errors and passing optional model reasoning effort.

    Usage details

    Who can use it: AI-Infra-Guard v4.6.10 users and scans using NVD-based AI product security data.

    How to use it: Update the AI-Infra-Guard repository to v4.6.10 or install the relevant release/setup package; check the Added and Fixed sections in the release changelog.

    Tip: Plan the CVE data update as a separate data synchronization from the scanning code update.

    ✓ Verified by official sourceSource: AI-Infra-Guard v4.6.10 changelog
  • CapabilityHy-Memory SDK

    Embedding services can return responses without an index

    The Hy-Memory embedding client accepts results when the provider response has no `index` field and rejects malformed index mappings once. This improves compatibility with OpenAI-compatible embedding services that do not send an index field.

    Usage details

    Who can use it: Hy-Memory SDK applications using OpenAI-compatible embedding providers.

    How to use it: Use the current Hy-Memory version without changing your existing embedding Base URL or model configuration; mappings are still validated when batch responses include an index.

    Tip: When testing a provider's batch response, verify the index-free success and malformed index-mapping cases separately.

  • Capabilityteamai-cli

    teamai uninstall now removes workspace leftovers it created

    Uninstall now cleans up machine-wide tool hooks that are no longer in use, workflow worktree setups, the Git exclude setting, and empty learning/report branches in self mode. Resources used by another installation or commit are preserved and identified in the preview.

    Usage details

    Who can use it: teamai-cli project and self-mode uninstall users; Codex, Pi, OpenCode, Kiro, and Antigravity integrations.

    How to use it: In the current teamai version, run `teamai uninstall` at the project root or `teamai uninstall --agent <agent>` for a specific tool; review preserved resources in the preview before proceeding.

    Tip: If other installations are present on the shared machine, read the preserved hooks in the uninstall output; use `teamai hooks remove` rather than forcibly deleting them.

    ✓ Verified by official sourceSource: teamai-cli: uninstall temizliği
  • CapabilityHy-Memory SDK

    Embedding retries prioritize HTTP status codes

    The embedding client prioritizes the HTTP status code over misleading words in the error message when deciding whether to retry. It retries 408, 429, and 5xx responses, while other statuses fail after a single attempt; transport connection errors continue to be retried.

    Usage details

    Who can use it: Hy-Memory SDK applications using OpenAI-compatible embedding services.

    How to use it: Set `max_retries` in the embedding configuration; the SDK automatically retries HTTP 408/429/5xx responses up to this limit.

    Tip: If your provider returns 429 or 5xx, increase `max_retries` and delay settings according to the provider's rate-limit policy.

  • CapabilityHy-Memory SDK

    Manual chat memory edits refresh modified time

    When a chat memory is edited manually, `gmt_modified` is now updated; creation and event times are preserved. In listings and statistics, the edited memory is sorted as the most recent item.

    Usage details

    Who can use it: Applications that update chat memories with HyMemoryClient; FAISS, Chroma, and Qdrant backends.

    How to use it: After calling `HyMemoryClient.update(memory_id, content, kind="chat")`, read the `updated_at` value from the results of `get`, `list_memories`, or `memory_stats`.

    Tip: Use the `gmt_modified`/`updated_at` field for synchronization; do not use `memory_at`, which marks when the event occurred, as the modification time.

  • CapabilityBrowserSkill

    BrowserSkill prevents blind mutation replay with Intent WAL

    The daemon and CLI add Intent WAL recording in under three milliseconds and bsk ledger protection. Blind mutation replay after an UNKNOWN status is blocked; `bsk sync` starts automatic reconciliation.

    Usage details

    Who can use it: Mutation automation using the BrowserSkill daemon/CLI; `$BSK_HOME/intent_ledger.jsonl`.

    How to use it: Use the normal mutation flow; for an uncertain result, run `bsk ledger status` and, if needed, `bsk ledger resolve`, then `bsk sync`.

    Tip: After an UNKNOWN result, do not manually repeat the same button or form action; reconcile the ledger status first.

  • FeatureBrowserSkill

    BrowserSkill Wave 2 introduces a zero-ceremony lifecycle and scope isolation

    The `bsk open/sync/close` lifecycle is now isolated by caller scope; aliases c1..cN and o1..oN, optional session/target-policy context, and guarded CDP soft-detach have been added. Soft-detach applies four safeguards in about 75 seconds while reducing close and tab-return delays to zero.

    Usage details

    Who can use it: Automations using the BrowserSkill daemon, CLI, and extension.

    How to use it: Open the session with `bsk open`, use `bsk sync` during work steps, and call `bsk close` when finished; specify the scope and target policy in the call context.

    Tip: For long-running tasks, call `sync` before every mutation; scope isolation prevents different callers from mixing up session state.

  • FeatureLoopForge Web

    Settings sidebar and capability/environment asset pages consolidated

    SOPs, runs, and scheduled tasks have moved to a 12-item settings sidebar. MCP, rules, and skills are tabbed under `Capability assets`; Dockerfiles, setup scripts, and sandboxes are tabbed under `Environment assets`; some legacy frontend modules have been removed.

    Usage details

    Who can use it: All LoopForge web interface users; visible pages depend on role and project permissions.

    How to use it: Open settings from the left navigation rail; select the relevant type tab on the `Capability assets` or `Environment assets` page. Legacy deep links redirect to the new consolidated views.

    Tip: Even if the MCP catalog is not visible, check the MCP tab and loading status on the assets page; the catalog remains available for general template installations.

  • CapabilityBrowserSkill

    BrowserSkill VOM v2 adds semantic selection and safe dropdown recovery

    VOM v2 supports affordance types, exact or case-insensitive selection by visible text, and offers suitable options for `option_ambiguous`/`option_not_found` responses. Recovery guidance for custom dropdowns and stealth injection cleanup have also been added.

    Usage details

    Who can use it: BrowserSkill virtual object model and form automations.

    How to use it: With VOM v2, provide the visible label for a select action; if the selection is ambiguous or not found, retry using the returned list of options.

    Tip: Prefer the label shown to users over the underlying value; if multiple options look the same, resolve the ambiguity instead of forcing the automation.

    ✓ Verified by official sourceSource: BrowserSkill: affordance-tipi VOM v2
  • SettingLoopForge

    Personal model catalogs arrive for Codex, Claude Code, and agent CLIs

    Personal settings now let users retrieve, select, reorder, and rename models for Codex, Claude Code, and protocol-based agent CLIs. GPT-6.1 Terra and Claude Haiku 5.5 have been added to the built-in catalogs.

    Usage details

    Who can use it: LoopForge personal model and CLI settings; accounts using official keys or a custom Base URL.

    How to use it: Open the relevant CLI model catalog in personal settings; retrieve models, select and reorder them, or edit their display names. For a custom Base URL, if the catalog is left empty, the service’s `/models` list is used.

    Tip: Move frequently used models to the top; on official endpoints, a saved personal catalog takes precedence over automatic discovery.

    ✓ Verified by official sourceModels: GPT-6.1 Terra, Claude Haiku 5.5Source: LoopForge: kişisel model kataloğu ve yeni yerleşik modeller
  • APIBrowserSkill

    FillResult carries actual and formatted value information

    The FillResult protocol adds the non-sensitive `actual_value` and `formatted` fields, along with the `uncertain_prior_effect` status. `ledger.request` is also a new protocol field for carrying the security and audit context of a mutation request.

    Usage details

    Who can use it: Clients using BrowserSkill protocol 1.9 and fill actions.

    How to use it: Read the `actual_value`, `formatted`, and `uncertain_prior_effect` fields in the fill response; support the protocol's `ledger.request` field in new client submissions.

    Tip: `actual_value` is abbreviated to avoid including sensitive information; use it as a verification hint in the audit record, not as the full form value.

  • CapabilityBrowserSkill

    Row key protection added for virtual scrolling

    BrowserSkill now verifies reused DOM rows in virtual lists against their business keys. If a row has been recycled, it stops the action with statuses such as `permission_denied` or `virtual_row_recycled` and provides a `cooperate sync` instruction.

    Usage details

    Who can use it: Automations for tables and lists with virtual scrolling.

    How to use it: For virtual list actions, provide the row’s business key or `rowAnchor` information; if an error occurs, run `cooperate sync` and resolve the row again.

    Tip: Don’t rely on the displayed row number; use the persistent business key as the anchor.

  • AvailabilityLoopForge

    Built-in `local` default sandbox available for SOP creation

    LoopForge now provides a built-in `local` sandbox profile in memory, even if the catalog SQL has not been installed. If a saved `local` profile exists, it takes precedence, and the default profile cannot be deleted.

    Usage details

    Who can use it: LoopForge instances installed without catalog data and users creating SOPs.

    How to use it: Select the `local` profile in the sandbox field when creating an SOP; for a custom driver or environment, edit the existing `local` profile or create a new one.

    Tip: When troubleshooting installation issues, first check that `/api/bootstrap` and the project sandbox list return the `local` profile instead of adding a sandbox row.

  • APIteamai-cli

    `teamai push --branch` reuses an existing local or remote branch

    If an explicit `--branch` target already exists locally, teamai checks it out and extends it; if it exists only on origin, teamai creates it from the origin tip. This prevents subsequent pushes to the same target from stopping with checkout or non-fast-forward errors.

    Usage details

    Who can use it: Team repositories using an explicit branch target with teamai push.

    How to use it: Run `teamai push --branch <ad>` again with the same target name; if the target exists locally or on origin, teamai reuses it.

    Tip: Don’t use the default branch as the target; choose an explicit review branch and check the status of any previous open PR.

  • APIteamai-cli

    Open-branch push safely merges teamai.yaml changes

    `teamai push --branch <ad>` now merges local `teamai.yaml` edits with changes on the target branch when the default branch is updated. Independent changes are preserved, genuine conflicts on the same key are reported as explicit errors, and CRLF-only differences are not counted as changes.

    Usage details

    Who can use it: Team configurations using `teamai push --branch`.

    How to use it: Run `teamai push --branch <ad>`; if a conflict occurs, manually reconcile the reported `teamai.yaml` keys and repeat the push to the same branch.

    Tip: Use an editor that preserves YAML comments and line endings; a CRLF-only change does not trigger a new push.

  • SettingLoopForge

    Requirement clarification is now optional

    The automatic score gate for requirements and the wait during the clarification stage have been removed. Users can start AI clarification if they want; an unclarified requirement no longer blocks SOP deployment.

    Usage details

    Who can use it: LoopForge requirements and SOP users; the new default behavior for all projects.

    How to use it: Create the requirement as usual and deploy the SOP directly. If AI clarification is needed, start the `AI ihtiyaç açıklaması` workflow as a user from the requirement panel.

    Tip: Skip the extra round and deploy directly when requirements are already clear enough; use clarification as an optional quality check for ambiguous work.

  • APIHy-Memory SDK

    Hy-Memory SDK can export and import memories with their IDs and vectors

    HyMemoryClient can now export a user's memories with their IDs and vectors and import them into another installation. This provides a new public SDK workflow for migrating and backing up user memories.

    Usage details

    Who can use it: Python applications using the Hy-Memory SDK; developers migrating user memories.

    How to use it: Call the SDK's memory export/import helpers in a HyMemoryClient setup; import the transfer data into the target client with the relevant user ID and verify the ID/vector mapping.

    Tip: Before importing, verify that the target embedding dimension is compatible with the source vectors.

  • Plugin / IntegrationLoopForge Local

    LoopForge adds Local, Hermes Agent, and DeepSeek Harness integrations

    LoopForge Local can now run inside DeepSeek Harness and Hermes Agent. Portable Edition adapters, installation instructions, and single-context fallback behavior when needed have been added for both environments.

    Usage details

    Who can use it: Portable Edition projects using DeepSeek Harness and Hermes Agent.

    How to use it: In the target project, run `npx -y loopforge-cli@latest skills install dsh` or `npx -y loopforge-cli@latest skills install hermes`. For Hermes, run `hermes skills trust .` in the project root and use `/devflow <request>`.

    Tip: In Hermes, you can also install the optional `local/hermes-plugin/` package for standalone task status and approval commands.

  • Plugin / IntegrationLoopForge

    ZCode added to LoopForge chat and SOP executors

    LoopForge has added ZCode as a new agent CLI executor for chat and SOP runs. If no model is specified at the default ZCode endpoint, GLM-5.3 is used; a model name is required for a custom endpoint.

    Usage details

    Who can use it: LoopForge chat, SOP, and executor users with a personal ZCode API key.

    How to use it: Add ZCode to agent CLI credentials in personal settings; select ZCode when choosing a chat or SOP executor. If you use a custom endpoint, specify the GLM-5.3 or GLM-5.3 Flash model name explicitly.

    Tip: Do not leave the model field blank for a custom endpoint; default model behavior applies only to the official endpoint.

    ✓ Verified by official sourceModels: GLM-5.3, GLM-5.3 FlashSource: LoopForge: ZCode agent CLI executor desteği
  • CapabilityLoopForge

    LoopForge can run SOP branches in parallel

    SOP branches that are ready and share a common predecessor now run together in the same workspace. If one branch is retried after cancellation, the canceled sibling branches are also reconciled correctly.

    Usage details

    Who can use it: LoopForge SOP runs and workflows that use parallel branching.

    How to use it: In the SOP design, create multiple successful branches from the same predecessor step and run them; the branches start in parallel automatically when ready.

    Tip: Define the merge step as dependent on all parallel branches; a fast successor does not need to wait for slower siblings.

    ✓ Verified by official sourceSource: LoopForge: paralel SOP dalları
  • SettingMemHub

    MemHub can select graph and vector backends via CLI or API

    In addition to the default Kuzu + Chroma pair, MemHub can now configure a Neo4j or Memgraph graph server and a Qdrant vector server. The settings are persisted as `storage_backends`; connections are tested before being saved, and credentials are stored in the secret store.

    Usage details

    Who can use it: MemHub operators; deployments using a Neo4j/Memgraph server or a Qdrant server.

    How to use it: Use the `memhub settings backends show|set|test|reset` commands, or call the `GET/PUT/DELETE /api/v1/settings/storage-backends` and `POST /api/v1/settings/storage-backends/test` endpoints.

    Tip: Note that switching backends does not migrate memory; existing Kuzu/Chroma data is retained and becomes visible again if you switch back.

  • FeatureMemHub

    MemHub backups now include memories and SDK stores

    `memhub backup create` and `POST /api/v1/backup` now back up not only the management SQLite database but also memories, raw turns, and local SDK stores; for Qdrant, they export memories as JSONL through the SDK. Schema 2 restore validates files, checks memory counts, and rolls back if inconsistencies are found.

    Usage details

    Who can use it: MemHub local Chroma/Kuzu or server-based Qdrant/Neo4j/Memgraph installations; v2 backup and restore flow.

    How to use it: Run `memhub backup create`; confirm with `--yes` after the size preview. For the API, use the `GET /api/v1/backup/estimate`, `POST /api/v1/backup`, and `POST /api/v1/backup/restore` endpoints.

    Tip: Check free disk space and the backend type before backing up; restoring to a different backend is intentionally rejected.

  • Plugin / Integrationteamai-cli

    teamai local-agent serves models to OpenClaw, Hermes, and DSH via MCP

    teamai can now write team MCP servers and models delivered over HTTP to OpenClaw, Hermes, and DeepSeek Harness user configurations. The pull, listing, doctor, and uninstall flows manage these integrations while preserving ownership records.

    Usage details

    Who can use it: teamai-cli workspaces using OpenClaw, Hermes, and DeepSeek Harness.

    How to use it: Define teamai MCP and model sources in the workspace; run `teamai pull` or the local agent flow. Verify the available sources with `teamai mcp list` and `teamai doctor`; uninstall with `teamai uninstall`.

    Tip: If multiple workspaces use the same MCP entry, check ownership with doctor before uninstalling; entries owned by others are preserved.

  • CapabilityLoopForge

    Scheduled AI tasks run a regular chat turn using the selected CLI

    AI scheduled tasks are no longer limited to CodeBuddy; they run as a regular chat turn on the selected supported CLI. The fixed five-minute tool gateway limit has been removed, and Codex and Claude Code use their own account keys.

    Usage details

    Who can use it: LoopForge AI scheduled tasks; accounts supported by CodeBuddy, Codex, and Claude Code.

    How to use it: When creating an AI scheduled task, select a CodeBuddy model or `codex:<model>` or `claude-code:<model>` in the `assistantModel` field; configure a personal key for the selected CLI.

    Tip: For long tasks, verify the selected CLI account key in advance; the shared CodeBuddy key is not passed to other CLIs.

  • CapabilityBrowserSkill

    BrowserSkill adds an unverified action mode to evaluations

    `bsk evaluate` can now evaluate an action flow with `--unverified` without comparing observations. The result is marked `verified:false`, and the mode cannot be used together with `--expect`, `--observation`, or `--depends-on`.

    Usage details

    Who can use it: BrowserSkill CLI and MCP evaluation workflows; protocol 1.9.

    How to use it: Use `bsk evaluate --unverified ...` or send `params.unverified` in an MCP request; use the same `unverified` option for DSH calls.

    Tip: Use this mode only when environmental observation verification is unavailable; do not confuse `verified:false` in reports with verified success.

  • Plugin / IntegrationBrowserSkill

    BrowserSkill adds skill installation targets for Qoder and Qoder CN

    BrowserSkill has released separate skill installation targets for Qoder and Qoder CN. Skills are installed in the relevant user directories and use the same CLI workflow as existing tool installations.

    Usage details

    Who can use it: BrowserSkill users who use Qoder and Qoder CN.

    How to use it: Run `bsk install-skill --harness qoder` or `bsk install-skill --harness qoder-cn`; installation goes to `~/.qoder/skills` or `~/.qoder-cn/skills`, respectively.

    Tip: When using Qoder CN, explicitly select the `qoder-cn` target; the general `qoder` target writes to a different user directory.

  • SettingDiagSafe

    DiagSafe adds an endpoint pool and a default output path

    DiagSafe can now use an endpoint pool with reasoning partitions and selects a default output location when no output path is specified. This makes it easier to manage tasks across different provider endpoints and find results in a consistent location.

    Usage details

    Who can use it: Tencent-Hunyuan DiagSafe CLI users.

    How to use it: Define the endpoint pool in the DiagSafe configuration; leaving the output parameter empty uses the default path. Check the detailed options with `diagsafe --help`.

    Tip: Specify the output path explicitly for production runs; use the default path for quick local tests.

  • CapabilityHy-Memory SDK

    Hy-Memory re-embed operations support Qdrant vector spaces

    The SDK's scan, diff, plan, reembed, verify, and drop workflows now work with a Qdrant server or local Qdrant space. The operation can be resumed, canceled at batch boundaries, and verified without changing the source collection, as with Chroma.

    Usage details

    Who can use it: Hy-Memory SDK applications using the Qdrant backend; source and target spaces in the same store.

    How to use it: In the Hy-Memory re-embed API, provide `provider="qdrant"`, host/port/api_key or persist_directory, and `prefix`; first run the scan/diff/plan workflow, then reembed and verify.

    Tip: Use a prefix for each home in Qdrant and create the target space on the same server; a different store for the source and target is rejected.

  • Plugin / IntegrationLoopForge Main Bot

    Lightweight GitHub tools added for Main Bot without opening a workspace

    LoopForge Main Bot can now use the GitHub tool catalog without opening a workspace. This adds GitHub information retrieval and related helper calls to the chat tool’s existing tool list.

    Usage details

    Who can use it: LoopForge Main Bot users; projects with GitHub tool access and authorization.

    How to use it: Open a Main Bot chat in the AI workspace and send a request for GitHub information; GitHub tools appear in the tool catalog without opening a workspace.

    Tip: Use the existing approval flow for write operations; this change makes the tools visible in the catalog but does not remove authorization checks.

  • FeatureLoopForge Chat

    Unread chat replies are marked with a blue dot in the sidebar

    Replies completed in another conversation are marked with a blue dot in the session list until the user opens that conversation. Read status is stored per user on the server, so it is preserved across devices.

    Usage details

    Who can use it: LoopForge chat users; conversations they participate in.

    How to use it: Open the chat sidebar; when you select a conversation marked with a blue dot, the reply is marked as read via the `/api/assistant/sessions/{id}/read` call.

    Tip: In workspaces that stay open for a long time, check the sidebar so you don’t miss replies completed in other sessions.

  • SettingLoopForge

    CLI and model selectors added to the SOP AI editing assistant

    The SOP AI editing assistant now lets you select a CLI and model from your personal model catalog. Changing the CLI starts a new chat but preserves the current drawing; changing the model keeps the session within the same CLI.

    Usage details

    Who can use it: AI editing assistant in the SOP design screen; configured personal CLI models.

    How to use it: Open the SOP design screen, open the AI editing assistant, and choose from the `CLI` and `Model` buttons below the message field; you can also search the lists.

    Tip: Preserve context by changing the model within the same CLI; switching between CLIs preserves the drawing but resets the chat context.

  • Settingteamai-cli

    teamai-cli can register installations and update other npm-global copies

    Each time it runs, `teamai update` registers the copy in `~/.teamai/installations.json`; `teamai update` and `teamai doctor` list these installations. With `updateOtherInstallations: true`, the current up-to-date copy can update another registered npm-global copy using its own `--prefix` field.

    Usage details

    Who can use it: Users running multiple teamai-cli copies on the same machine; automatic update policy.

    How to use it: Add `updateOtherInstallations: true` to `~/.teamai/config.yaml`, then run `teamai update` or `teamai doctor` to check installations.

    Tip: Volta, vendored, and local checkout copies are not included in automatic npm updates; follow the manual instructions in the doctor output.

  • APIteamai-cli

    GitHub PR REST fallback now also uses GH_TOKEN

    When GitHub CLI is unavailable, teamai's PR REST fallback now checks `GITHUB_TOKEN` or `GH_TOKEN` using shared token resolution logic. This lets environments with only GH_TOKEN configured retrieve PR data, commits, and diffs.

    Usage details

    Who can use it: teamai users using the GitHub PR provider in CI or local environments where the `gh` command is unavailable.

    How to use it: Set `GH_TOKEN` or `GITHUB_TOKEN` in the environment; the REST fallback is used automatically when `teamai` runs a PR-reading or workflow flow.

    Tip: If you set both variables, use teamai's shared token precedence; do not put the token in team files.

  • FeatureLoopForge Quick Access

    Quick access wizard explains repository linking permissions earlier

    When a non-Owner member in a shared project tries to link a GitHub repository, the warning now appears at the project step rather than the final step. The wizard offers the options to create and link their own project, or skip linking and continue.

    Usage details

    Who can use it: LoopForge quick access wizard; non-Owner members in shared projects.

    How to use it: In Quick Access, open the `Project and repository` step; in the permission warning, choose `Create and link my own project` or `Skip linking and continue`.

    Tip: If you are not the owner of the current project, choose the new project option directly to avoid losing the repository URL.

  • APILoopForge Issues

    Undeletable requirements now show the reason and related run information

    List, detail, and DELETE conflict responses now include the deletion block's code, a human-readable reason, and the related run or task ID. Empty time fields for stages that haven't started are also no longer shown in JSON.

    Usage details

    Who can use it: LoopForge requirement lists, detail screen, and API consumers.

    How to use it: Read the `deletionBlock` field in the list or detail response; if DELETE returns 409, use the same field and the reason in the `error` text.

    Tip: Show the `deletionBlock.code`, `reason`, `runId`, and `taskId` fields on the client side instead of mapping fixed error messages.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 16

  • AgentAgentScope 2.0

    AgentScope 2.0 highlights HarnessAgent as the entry point for long-running tasks

    AgentScope 2.0 brings together task planning, persistent state, workspace, context compression, subagents, memory, permissions, and sandbox execution as a general-purpose Agent Harness. HarnessAgent is presented as the primary development entry point for long-horizon enterprise tasks, separating server-side orchestration of model/tool calls from isolated tool and code execution.

    Usage details

    Who can use it: Apache-2.0-licensed AgentScope 2.0; requires Python 3.11 or later and a local, Docker, K8s, or other supported sandbox execution environment.

    How to use it: In a Python 3.11+ environment, run `uv pip install agentscope`; in the HarnessAgent configuration in the AgentScope documentation, enable the model, workspace, task list, plan mode, and execution environment.

    Tip: For long-running tasks, keep the goal, progress, and evidence in separate state fields; do not treat constraints suggested by the model as authorized decisions without validating them in the application.

  • FeatureAlibaba Cloud DevOps AI Assistant

    Alibaba Cloud DevOps AI Assistant expands lifecycle automation

    The November 11 update embeds the assistant in requirements, testing, code, pipeline, and release pages. New workflows cover automatic test case generation, AI Smart Fix for scan findings, natural-language pipeline generation and troubleshooting, release risk assessment, and scheduled or event-triggered tasks; task results can be sent to email, DingTalk, or Feishu.

    Usage details

    Who can use it: For Alibaba Cloud DevOps users, via the AI Assistant icon at the top of the page or the relevant buttons on requirements, code, pipeline, and personal workbench pages.

    How to use it: Sign in to Alibaba Cloud DevOps and open the AI Assistant icon at the top; to automate tasks, create a scheduled or event-triggered task in Personal Workbench → AI Assistant Tasks.

    Tip: Do not merge branch and merge request changes from AI Smart Fix before human review and validation of the basic test results.

  • Pricing / PlanToken Plan

    Qwen Cloud launches Token Plan Individual and lowers the price of the Team plan

    QwenCloud’s upgraded Token Plan page announces the launch of the Individual plan and a more affordable Team plan. The page mentions an approximately 40 percent discount, access to text, image, speech, and video generation models under a single plan, and tool integrations.

    Usage details

    Who can use it: Monthly, quarterly, or annual subscription options for Individual and Team plans; the page lists tools such as Qwen Code, Cline, Claude Code, Cursor, OpenCode, Codex, Kilo CLI, and OpenClaw.

    How to use it: Choose a tier and subscribe on the Token Plan page; then get the API key specific to your plan and connect using Token Plan credentials in the provider settings of a supported tool.

    Tip: Save the API key in a secure place when you create it; according to the Token Plan documentation, the key cannot be viewed again later.

    ✓ Verified by official sourceModels: Qwen4-Max-Preview, Qwen3.8-MaxSource: Token Plan | QwenCloudSource: Alibaba Cloud Model Studio: Token Plan overview
  • AvailabilityQwen Code

    Qwen Code 1.0 Preview combines Qoder capabilities and native packages

    Qwen Code 1.0 Preview combines Qoder and Qwen Code’s code understanding, implementation, debugging, and testing agent capabilities in a single preview. It supports Qwen and Qoder accounts, multiple model providers, compatible custom endpoints, session resumption, and non-interactive execution.

    Usage details

    Who can use it: Native packages for macOS, Linux, and Windows that do not require a Node.js installation; the release is a preview, and the interface, configuration, and behavior may change.

    How to use it: Download the v1.0.0-preview package for your operating system and architecture from the GitHub release page; sign in with your existing account or configure a compatible provider/custom endpoint.

    Tip: Try the preview package in a separate environment before using it in production automation; validate session resumption and non-interactive workflows with your existing scripts.

    ✓ Verified by official sourceSource: Qwen Code 1.0 Preview release
  • APIAlibaba Cloud Model Studio API

    Output budget in the Qwen4-Max-Preview API limits thinking and the answer together

    For OpenAI Chat Completions calls to Qwen4-Max-Preview, max_tokens and max_completion_tokens have been made equivalent. This limit covers the combined total of thinking tokens and final answer tokens; the behavior of existing Qwen models is unchanged.

    Usage details

    Who can use it: Only in the Token Plan preview for Qwen4-Max-Preview and on the OpenAI Chat Completions-compatible endpoint.

    How to use it: Set either max_tokens or max_completion_tokens in the request, and choose a value sufficient for the combined total of thinking and the final answer; do not use both at the same time.

    Tip: Leave extra headroom for the xhigh thinking default, even if you expect a short answer; otherwise, generation may hit the budget before the answer is complete.

    ✓ Verified by official sourceModels: Qwen4-Max-PreviewSource: Alibaba Cloud Model Studio: qwen4-max-preview kullanım notları
  • ModelQwen4-Max-Preview

    Qwen4-Max-Preview opens for preview access on Token Plan and selected apps

    Alibaba released Qwen4-Max-Preview, the first accessible model in the Qwen4 family. The model supports text generation, image and video understanding, deep thinking, tool calling, web search, and structured outputs; it has a context length of 1 million tokens and supports compatibility with OpenAI Chat Completions, Responses, and Anthropic Messages.

    Usage details

    Who can use it: Available only through Token Plan; the preview runs from 2026-11-11 10:00 to 2026-11-25 24:00 UTC+8. According to Qwen’s announcement, early access is also available in QwenWork, Qoder, Qwen App, and QwenStudio; the model will be retired when the preview ends.

    How to use it: Subscribe to Token Plan Personal or Team, get the API key specific to your plan, use https://token-plan.ap-southeast-1.maas.aliyuncs.com as the Base URL, and set the model ID to qwen4-max-preview.

    Tip: Use the 1 million-token context for long repo or document tasks; because thinking and the final answer share the same output budget, do not set the budget based only on answer length.

  • AgentAlibaba Cloud AgentCore

    AgentCore defines Team/Task collaboration for people and specialist agents

    AgentCore’s November 10 announcement describes joint work by people and specialist Agents through the AgentIdentity, Matrix, Team, Task, SubTask, and Result objects. TeamHarness provides a task CLI based on Leader/Worker roles, while AgentLoop observes tasks, subtasks, traces, and deliverables and ties them to evaluation.

    Usage details

    Who can use it: For enterprise AgentCore use; access options via Web, DingTalk, Feishu, WeCom, and AG-UI are listed.

    How to use it: Define the human Admin, Leader Agent, and Worker Agent roles in AgentCore; set up task and subtask workflows according to the Team/Task documentation, then launch them through the selected Web or enterprise messaging channel.

    Tip: Define data and tool permissions separately for each Worker; verify that joining the same room does not automatically grant access to all resources.

  • AgentHiRoute

    HiRoute can route tasks by assigning plans to specific execution Agents

    Separate from model routing, HiRoute introduced task routing: plans can be assigned to execution Agents such as Codex CLI, Claude Code, Qoder CLI, Pi, DeepSeek Harness, or OpenCode. Hermes can route both models and tasks; OpenCode gains a dual role, able to delegate tasks and act as an execution Agent.

    Usage details

    Who can use it: For desktop, Web UI, or CLI setups with a HiRoute plan and locally installed execution Agents.

    How to use it: Assign a routing plan to an execution Agent; to submit a task, use `hiroute worker exec --plan <PLAN_ID> --cwd <WORKDIR> --no-wait -- "<task>"`, and to check its status, use `hiroute worker status --run <RUN_ID>`.

    Tip: Set up separate plans and Agents for information gathering, coding, and review; use --no-wait to run long jobs without blocking the main conversation.

  • AvailabilityHiRoute

    HiRoute makes its Windows desktop app and Linux browser interface available

    The November 10 announcement describes HiRoute’s Windows x64 desktop release and a Linux workflow for opening the same interface as the desktop app in a browser. The Windows package includes the command-line tool and background service; the Linux installation can be placed in the user’s home directory.

    Usage details

    Who can use it: Windows x64 desktop; browser-based Web UI on Linux x86_64 and ARM64. macOS support also remains available for macOS 15 and later.

    How to use it: Download and install the Windows package; on Linux, run `curl -fsSL https://hiroute.ai/install.sh | sh -s -- --console-mode loopback --console-port 5838`, followed by `hiroute service start --output json` and `hiroute console open`.

    Tip: Instead of installing a desktop environment on a Linux server, expose the loopback console through an SSH tunnel or behind a secure gateway.

  • Plugin / IntegrationHiRoute

    HiRoute combines subscription and API sources in the same model route

    When connecting a Codex or Claude subscription, you can also add the corresponding OpenAI or Anthropic API key and import the models in one go. Models with the same name appear as a single candidate; HiRoute can use the subscription first and fall back to the API when it is temporarily unavailable.

    Usage details

    Who can use it: On HiRoute installations using eligible ChatGPT Pro/Business and Claude Max/Ultra/Team subscriptions and their corresponding API connections.

    How to use it: Add the API key corresponding to the subscription connection, select and save the models, and set the combined subscription/API model as a route candidate in the model details.

    Tip: To avoid an unexpected bill when subscription credits run out, set a separate budget limit on the API key.

  • Pricing / PlanHiRoute

    HiRoute adds one-time and recurring spending budgets for API keys

    HiRoute now lets you set a one-time budget for an API key, a budget that renews on a schedule, or a shared budget for multiple keys. The amount is reserved before a request and reconciled when it completes; the budget is based on estimated traffic routed through HiRoute.

    Usage details

    Who can use it: For API keys connected to HiRoute and decision/normal model connections; requests with unknown prices may proceed without pricing.

    How to use it: In the key settings for the relevant API connection, select a budget type, set the amount or renewal schedule, and link the key to the relevant route.

    Tip: To avoid exceeding your monthly provider credits, set the renewing budget close to the credit amount, with a small safety margin.

  • SettingHiRoute

    HiRoute offers separate proxy and bypass settings for each API connection

    Each API connection can now connect directly, inherit the background service’s proxy environment, or use manual HTTP(S) proxy and bypass rules. Decision models sharing the same connection also follow its proxy settings.

    Usage details

    Who can use it: For API and decision model connections in HiRoute; for desktop, Web UI, and CLI use.

    How to use it: On the connection editing screen, select Direct, inherit, or Manual for the proxy mode; save the manual HTTP(S) address and bypass rules, then retry the connection.

    Tip: Add internal network endpoints to the bypass list; routing all traffic through the proxy can cause latency and authentication issues.

  • FeatureHiRoute

    HiRoute adds a trust level and performance report for new models

    HiRoute’s November 10 announcement introduces a New Models feed on the home page to test a new model against the selected primary model in a controlled way. The trust level can be adjusted in increments; the performance report recommends raising or lowering it based on stage scores and cost estimates, and replays records before publishing the change.

    Usage details

    Who can use it: In local Agent routing flows covered by HiRoute 0.6.0; trials and reports are managed from the home page.

    How to use it: Open New Models on the HiRoute home page, select the new model and primary model, and assign them to Agent; set the trust level, review the report, and publish the recommended change.

    Tip: Assign the new model to low-difficulty steps first; review real outputs from at least a few different conversations before increasing the trust level.

  • Plugin / IntegrationHiRoute

    HiRoute enables three-question decision routing with System One and OpenAI Decisions

    HiRoute can select a model group by separately evaluating the previous step, the difficulty of the current step, and whether the previous question carried over to the next step on each turn. The announcement also expands OpenAI Decisions connectivity and decision model connections compatible with Alibaba Cloud Model Studio, OpenRouter Jev, and TypeSafe.

    Usage details

    Who can use it: On local installations that use HiRoute decision model settings; OpenAI, OpenRouter, and Alibaba Cloud Model Studio connections can reuse existing API keys.

    How to use it: Add a decision provider under Models → Decision Models, set the context window limit, and publish a routing plan that uses System One.

    Tip: Track the decision model’s cost and latency separately; the three-question evaluation improves routing decisions but does not guarantee task completion on its own.

    ✓ Verified by official sourceModels: Jev 1.14, OpenAI DecisionsSource: HiRoute New Release: Choose the Best-Fit Model for Your Real-World Work
  • AvailabilityHiRoute

    HiRoute updates its model catalog and new model recommendations independently of the app

    Official model information, context windows, and new model recommendations on the home page can now be published through a separate HiRoute model catalog without waiting for an app release. This reduces the need to reinstall the app to get model metadata.

    Usage details

    Who can use it: Desktop, Web UI, and CLI clients that use HiRoute’s model catalog; catalog connectivity and update behavior depend on the installation’s network settings.

    How to use it: Make sure the HiRoute model catalog setting is enabled, refresh the catalog, and check the current context window/recommendation information in the model details.

    Tip: If you use your own endpoint, compare the model ID and context limit in the catalog with the actual provider behavior.

  • SettingHiRoute

    HiRoute compresses local records and adds a default 20 GB limit

    Full request snapshots and protocol logs are no longer stored repeatedly during long conversations; content needed to show history, scores, and the model used is retained. Local records use up to 20 GB by default; this limit can be changed or disabled, while the seven-day retention period remains in place.

    Usage details

    Who can use it: On all platforms that use HiRoute’s local records and conversation history.

    How to use it: In HiRoute’s record/storage settings, leave the local record budget at 20 GB, set a different upper limit, or disable the limit; also keep the seven-day retention rule in mind.

    Tip: Set up disk alerts on machines that run many long tasks; disabling the limit makes sense only if you have sufficient storage monitoring.

Watchlist

Muse Wonder

44
Meta · Meta AI

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

Claude Haiku 5.5

36
Anthropic · Claude

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

Muse Ember

36
Meta · Meta AI

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

Grok 4.7

35
xAI · Grok

The current model in the Grok app and API. Elon Musk says pretraining for Grok 4.8 is finished; it will be made available to a limited group of Grok Bot users this week. There’s no API ID or pricing for Grok 4.8 yet.

Tencent Hy4

35
Tencent · Hy

Available through the API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, free on WorkBuddy and CodeBuddy through November 15. Today’s Tencent developments are in agents and memory tools, not models.

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba’s most highly rated model. Qwen4-Max-Preview launched on November 11 (1M context, available only through Token Plan, preview through November 25); it doesn’t yet have an Artificial Analysis score.

Sources