YZ·RADARyz.mikail.net
Daily edition · 7 November 2026

An agent command center for Codex, a security patch for Claude Code, and Spaces for Muse

No changes to the TOP 10 for a third day: all ten models remain in the same order with the same scores as yesterday. Gemini 4 Carbon leads with 60 points, followed by Claude Opus 5.5 (59) and GPT-6 Astra (57); the Artificial Analysis index is at v5.0.1. The latest accessible view of the LMArena text ranking is still dated October 27, and Gemini 4 Carbon is also in first place there, so we treat it as a cautious second opinion. The watchlist is unchanged too. Two small notes: Databricks retired the older DeepSeek V4 Flash (0731) on its platform and directed users to V4.1 Flash; the report that Tencent increased Hy4 inference capacity tenfold is, for now, sourced only to TechNode.

Today's focus is developer tools. OpenAI Codex CLI 0.167.0 adds an agent command center (`codex agents`) for monitoring multiple agents from a single screen, alternative sign-in with Google Cloud Vertex AI and an API key, macOS/Windows system proxy support, and preservation of saved chats when switching providers; the Python, Node, Go, and Ruby SDKs also recognize GPT-6 Rosalind and GIF/SVG image output. On the Anthropic side, Claude Code 2.1.319 is an important security release: permission-check bypasses via `sed` and Bash environment prefixes, as well as sandbox escapes during worktree creation, have been closed. The same version runs MCP tasks in the background and adds the `permissions.denyInsteadOfPrompting` setting, which rejects calls instead of showing a permission prompt. If you use Claude Code, update today. In Enterprise, adding cost-code tags to projects and tracking spend is now in beta; Gateway adds periodic spend views and budget-priority rules.

Google expanded its enterprise data and agent infrastructure. Gemini Enterprise now has Data Cloud connectors, enabling natural-language queries of live data in BigQuery, Spanner, Cloud SQL, and AlloyDB using user permissions; Looker is in preview. Agent Platform now has a dedicated Memory Bank console page, and Agent Registry skills can be scanned for data-leakage and prompt-injection risks. Gemini 4 Argon can generate up to 1 million tokens of output in a single request through the Interactions API. Requests to `gemini-embedding-2-preview` now route to the GA model; reindexing is not required. Administrators should note two dates: Gemini Notebook Enterprise addresses switch to the `notebook` domain on November 16 (update allowlists), and the bundled Code Assist entitlement ends on March 31, 2027. Meta added Spaces to Muse, turning an idea into a persistent tracker or planner; specialist models trained with Muse Wonder are also being deployed. According to Meta, these models forget critical information 2.4 times less often.

On the China front, Qwen Code v0.26.9 adds installation of large Skill ZIPs through the daemon and Web Shell, persistent multiple-choice questions in Managed Agent sessions, and automation enabled by default. Agent costs can now be tracked by session and turn in Alibaba Cloud AI Gateway. Kimi Code 2.5.2 fixes MCP and steer bugs; Microsoft announced that Kimi K3.1 Pro is coming to the Foundry catalog. Z.ai's AutoClaw gives accounts that sign in from November 9–15 a one-time 100 million tokens. Tencent released LoopForge v0.3, an agent platform that runs on your own server, and WorkRally CLI 2.12, which adds enterprise knowledge-base commands. There were no new user-facing changes at xAI or Xiaomi; DeepSeek only saw the model retirement on Databricks. A total of 96 new developments were added.

Today’s TOP 10

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

  1. 01

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    At the top with 60 points. Access is still limited to selected cyber defenders and Fairwind partners; general API and app access is not yet available.Released: 28 October 2026

  2. 02

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Anthropic's flagship model; available on Claude Pro/Max/Team/Enterprise and via API. Project-tag cost tracking is now in beta for Enterprise; Claude Code 2.1.319 closes permission and sandbox vulnerabilities.Released: 22 September 2026

  3. 03

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

    OpenAI's most powerful GPT-6 model, for long and challenging reasoning tasks. Codex CLI 0.167.0 adds an agent command center and sign-in through Vertex AI.Released: 3 September 2026

  4. 04

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

    The default model on paid ChatGPT plans; API pricing is $2 / $10.Released: 29 September 2026

  5. 05

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

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

  6. 06

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

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

  7. 07

    Gemini 4 Argon

    Google · Gemini=
    50pts

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

  8. 08

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta's most powerful Muse model; no public API. Muse adds Spaces and specialist models trained with Muse Wonder (Active Memory, Active Focus).Released: 12 October 2026

  9. 09

    Claude Fable 5.1

    Anthropic · Claude=
    50pts

    The model with the highest accuracy; at $10 / $50, it is the most expensive on the list.Released: 1 September 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). According to Microsoft's announcement, it is coming to the Foundry catalog ($2.30 / $11.50).Released: 23 October 2026

Key signals

  • The TOP 10's order and scores are unchanged for a third day: Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57. The watchlist is unchanged; 96 new developments were added today.
  • OpenAI: Codex CLI 0.167.0 adds an agent command center (`codex agents`), sign-in with Vertex AI/API key, and system proxy support. The Python/Node/Go/Ruby SDKs recognize GPT-6 Rosalind and GIF/SVG image output.
  • Anthropic: Claude Code 2.1.319 closes permission bypasses and a worktree sandbox escape; MCP tasks run in the background, and the `denyInsteadOfPrompting` setting is available. Project-tag cost tracking is in beta for Enterprise.
  • Google: Gemini Enterprise adds Data Cloud connectors, a Memory Bank console, and Agent Registry security scanning. Gemini 4 Argon supports 1M-token long output; Embedding 2 Preview routes to GA; Notebook addresses change on November 16.
  • Meta: Muse adds Spaces (from idea to personal dashboard and planner) and specialist models trained with Muse Wonder (Active Memory, Active Focus). Quality-based business searches are being enabled automatically in WhatsApp.
  • Qwen Code v0.26.9 (Skill ZIP installation, structured questions), session-based FinOps in AI Gateway; Kimi K3.1 Pro is coming to Foundry; AutoClaw gives 100M tokens November 9–15; Tencent LoopForge v0.3. No new changes at xAI or Xiaomi.

Worth trying today

01

Manage parallel agents from one screen in Codex

Update Codex CLI with `npm install -g @openai/codex@0.167.0` and open the agent command center with `codex agents --no-daemon`. You can see sessions running simultaneously in a single list and tell which ones have finished from unread-result indicators. If you're behind a corporate network, enable the system proxy option from the `/experimental` menu; if you use Google Cloud, connect to Vertex AI via `Sign in another way` on the sign-in screen.

OpenAI · ChatGPT
02

Keep unattended Claude Code jobs from getting stuck

First, update Claude Code to 2.1.319; this version closes permission and sandbox vulnerabilities. For overnight or CI jobs, add `permissions.denyInsteadOfPrompting: true` to your settings file: tool calls that would wait for permission are rejected, and the turn continues without waiting. `claude -p` now returns 1 for rejected actions when it cannot make a security decision, so check the exit code in your scripts. Use `--trust-workspace false` in untrusted repositories.

Anthropic · Claude
03

Set up a Space in Muse for recurring tasks

Open Muse and describe the experience you want in natural language—for example, “a dashboard that tracks my weekly training sessions, participant counts, and preparation steps.” Spaces turns this into a persistent tracker or planner. You can pin the Space to the sidebar or main tabs and schedule it to appear at specific times. A short description focused on a single goal works better; connect only the apps you want to keep up to date.

Meta · Meta AI
04

Complete the Embedding and Notebook migrations today

If you're using Gemini Embedding 2 Preview, change the model name in your code to `gemini-embedding-2`; the schema, pricing, and vector space are unchanged, so reindexing is not required. If you're a Gemini Enterprise administrator, add `notebook.cloud.google.com/*` to your firewall and proxy allowlists before November 16. The bundled Code Assist entitlement ends on March 31, 2027; start planning your move to Antigravity.

Google · Gemini

This edition’s updates by AI

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

Google · Gemini

#1 Gemini 4 Carbon#7 Gemini 4 Argon
Full history (244) →

New in this edition · 10

  • FeatureGemini Enterprise Agent Platform

    Dedicated Cloud Console management arrives for Agent Platform Memory Bank

    A separate page for Memory Bank is now available in Cloud Console: create and manage independent memory instances, inspect structured and unstructured memories, filter by scope, and view revision history—all in one place. Generation, retrieval, retention, schema, and ADK connection code settings can also be managed from the console.

    Usage details

    Who can use it: Cloud projects using Gemini Enterprise Agent Platform Memory Bank; memory management in Agent Runtime deployments has also been updated.

    How to use it: In Google Cloud Console, open Agent Platform > Memory Bank and create an instance with Create memory bank. Set scope, schema, retrieval, and retention options in the Memories and Configuration tabs, then get the ADK code with Link to an agent.

    Tip: Use separate scopes for user or agent identities, and set TTL/retention up front to automatically clean up old memories.

  • FeatureAgent Registry

    Security scanning and organization policies arrive for Agent Registry skills

    Agent Registry can now automatically scan standalone skill revision packages for risks such as data exfiltration, exposed credentials, and prompt injection. Finding sources can be inspected, scanning policy can be applied across an organization, and CMEK support for skill payloads is available as Preview.

    Usage details

    Who can use it: Google Cloud Agent Registry users; the default policy is `ENABLED_BLOCKING`, while custom Organization Policy constraints at the organization or folder level are optional.

    How to use it: In the console, choose a policy under Agent Registry > Skills > Add/Edit > Security scanning, or use the `gcloud agent-registry skills create SKILL_ID --scanning-policy=enabled-blocking` command; then inspect the revision scanResult and Finding resources.

    Tip: Because LLM-based scans can produce false positives and negatives, verify findings against file lines and behavior; keep the blocking policy in production.

  • RetiredGemini API

    Gemini Embedding 2 Preview requests automatically route to the GA model

    Google has deprecated the `gemini-embedding-2-preview` endpoint. Starting November 6, 2026, requests are automatically routed to the `gemini-embedding-2` model with the same schema, pricing, and embedding space; reindexing is not required.

    Usage details

    Who can use it: Projects using the Gemini API; requests to the old preview model ID automatically go to the GA model.

    How to use it: In your code, replace the model string `gemini-embedding-2-preview` with `gemini-embedding-2`; you do not need to change the API schema or vector indexes.

    Tip: Update the model ID now rather than relying on automatic routing, and monitor post-migration usage in the billing dashboard.

    ✓ Verified by official sourceModels: gemini-embedding-2-preview, gemini-embedding-2Source: Gemini API sürüm notlarıSource: Gemini API kullanımdan kaldırma takvimi
  • RetiredGemini Enterprise

    Code Assist entitlement in Gemini Enterprise ends on March 31, 2027

    The bundled Gemini Code Assist entitlement in existing Gemini Enterprise Standard and Plus subscriptions will be removed on March 31, 2027. Until then, support is limited to the Gemini 3.x family; new Gemini Enterprise subscriptions have not included this entitlement since October 9, 2026.

    Usage details

    Who can use it: Existing Gemini Enterprise Standard and Plus subscriptions; after the transition, coding support is provided through Antigravity, included with Gemini Enterprise editions.

    How to use it: As an admin, verify the end date of the Code Assist entitlement in your subscription and use the Antigravity IDE/CLI migration guide to move developer workflows before March 31, 2027.

    Tip: Inventory plugins and CI workflows tied to Gemini 3.x now, and validate them with a small pilot in Antigravity rather than waiting until the transition.

  • Plugin / IntegrationGemini Enterprise

    Live data queries enabled in Gemini Enterprise with Data Cloud connectors

    Data Cloud connectors let Gemini Enterprise query live enterprise data in natural language. BigQuery, Spanner, Cloud SQL, and AlloyDB are available in federated query mode; Looker also works as Preview with Knowledge Catalog and user permissions.

    Usage details

    Who can use it: Public Preview in Gemini Enterprise; BigQuery, Spanner, Cloud SQL, and AlloyDB federated query connectors, as well as Looker Preview, are limited by user permissions and row/column policies.

    How to use it: In the admin console, add and authorize the relevant Data Cloud data store/connector for the Gemini Enterprise app; users can then enable the connector in the Assistant tab and ask BigQuery or the database questions in natural language.

    Tip: Start with a read-only connection and enrich Knowledge Catalog business context with validated queries; set a default billing project to manage query costs.

  • OtherGemini Notebook Enterprise

    Gemini Notebook Enterprise URLs are moving to the `/go/` path and the `notebook` domain

    The `go/` path is live in global and US regions, rolling out in the EU, and will begin in other locations on November 9. Starting November 16, the `notebooklm` hostname will become `notebook`; existing links will redirect automatically, but firewall and proxy allowlists should be updated.

    Usage details

    Who can use it: Notebook in Gemini Notebook Enterprise and Gemini Enterprise; global, us, eu, and other listed regions.

    How to use it: Add `notebooklm.cloud.google.com/*` and `notebook.cloud.google.com/*` to your allowlist, along with the relevant `.cloud.google` domains if you use federation; for Notebook in Gemini Enterprise, also allow the `vertexaisearch.cloud.google.com/notebooklm/*` path.

    Tip: To avoid having to repeat the process for future path changes, allowlist the entire domains with `/*`, not just the old and new individual paths.

  • AvailabilityAgent Registry

    Agent Registry standalone skills reach general availability

    Standalone skills can now be registered, versioned, searched, and managed in Agent Registry as GA. `Skill`, immutable `SkillRevision` packages, and verified `Publisher` details can be used across agents with centralized governance.

    Usage details

    Who can use it: GA in Google Cloud Agent Registry; users can manage skills at the project and location scope.

    How to use it: Set up Agent Registry, then register a `SKILL.md` package in a ZIP using Skills > Add skill in the console or the `gcloud agent-registry skills create` command; then manage the revision and default revision lifecycle.

    Tip: Publish skill packages as immutable revisions and update the default revision in a controlled way; retain previous revisions for rollback.

  • SettingGemini Enterprise

    Gemini Enterprise model availability options now follow the resource location policy

    If the project’s resource locations organization policy does not allow the `global` location, availability options are disabled for Gemini models that are unsupported in the app’s location and send traffic to the global endpoint. If a model is already enabled, the admin is shown a warning.

    Usage details

    Who can use it: Gemini Enterprise admins applying a resource location organization policy, and models routed to the global endpoint.

    How to use it: In the Google Cloud console, open Gemini Enterprise > Feature Management > Model availability; check which model toggles are locked or display warnings due to policy, and enable only models supported in permitted locations.

    Tip: Check data residency and machine learning processing locations before enabling a model; do not assume global routing can bypass policy restrictions.

    ✓ Verified by official sourceSource: Google Cloud sürüm notları
  • Plugin / IntegrationGemini Enterprise

    Adobe Creative Cloud data store preview arrives in Gemini Enterprise

    The Adobe data store is now available as Public Preview. Users can search, share, and manage Creative Cloud assets in natural language through Gemini Enterprise; customize Adobe Express templates, apply Lightroom presets, and make Photoshop-style image edits.

    Usage details

    Who can use it: Available in Gemini Enterprise Standard, Plus, Pay-as-you-go, and Frontline editions; the data store is Public Preview, limited to global, us, and eu locations.

    How to use it: Select Adobe in the data store setup in the Gemini Enterprise admin console and authorize the connection; then use natural language in Gemini to request asset searches, sharing, or edits.

    Tip: Since this is a preview, start with a single Adobe data store, grant action permissions using least privilege, and review VPC Service Controls limitations in advance.

  • APIInteractions API

    Interactions API offers long decode of up to 1 million tokens for Gemini 4 Argon

    Long decode for up to 1,048,576 tokens in a single interaction is now GA in the Interactions API for Gemini 4 Argon. The server automatically continues and combines 65,536-token chunks in the background; processing can take up to 24 hours with `background=true`.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform Interactions API v1beta1; `gemini-4-argon` and `gemini-4-argon-cyber-permissive` models.

    How to use it: In an `interactions.create` call, set `background=true` and `max_output_tokens` to a value greater than 65,536; poll or stream the returned interaction ID and combine all `model_output` steps.

    Tip: Use this feature only for lengthy math, science, coding, or agent tasks; account for charges per hop and delays of several minutes.

    ✓ Verified by official sourceModels: Gemini 4 Argon, gemini-4-argon, gemini-4-argon-cyber-permissiveSource: Gemini Enterprise Agent Platform sürüm notlarıSource: Gemini API long decode kılavuzu

Anthropic · Claude

#2 Claude Opus 5.5#5 Claude Sonnet 5.5#9 Claude Fable 5.1Claude Haiku 5.5
Full history (836) →

New in this edition · 20

  • CapabilityClaude Code

    Bypasses in tool permission checks have been closed

    Fixed issues where `sed` calls could bypass checks in accept-edits and Auto mode, a Bash environment prefix could hide part of a command, and files could be changed via `/rewind` through a directory symlink.

    Usage details

    Who can use it: Claude Code 2.1.319.

    How to use it: Update, then review your existing permission and deny rules.

    Tip: Narrow permission rules based not only on the command name but also on the working directory and target path.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingClaude Code

    Auto mode behaves more cautiously when server checks are unavailable

    If telemetry or feature flag access is disabled and Claude Code cannot verify the server security check, it now asks the user or rejects the call instead of silently falling back to the built-in classifier. Behavior behind LLM gateways has also been clarified.

    Usage details

    Who can use it: Claude Code 2.1.319; Bedrock, Vertex, and Foundry are unaffected by this change.

    How to use it: Use `CLAUDE_CODE_AUTO_MODE_SERVER=0` for the previous built-in behavior; use `CLAUDE_CODE_AUTO_MODE_SERVER=1` to require server checks behind a gateway.

    Tip: If your organization settings include `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS`, explicitly set `CLAUDE_CODE_AUTO_MODE_SERVER=1` to enable server checks.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • APIClaude Code

    When there’s no security decision, claude -p returns exit code 1

    When an Auto mode action is rejected because there’s no security decision, `claude -p` now returns 1 instead of incorrectly returning 0 in both text and JSON output formats.

    Usage details

    Who can use it: Claude Code 2.1.319; headless and automated runs.

    How to use it: Check the exit code of your `claude -p` call; route status 1 to your error flow.

    Tip: In CI scripts, make decisions based on the exit code, not just the text output.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • Plugin / IntegrationClaude Code

    MCP tasks can run in the background in the terminal

    When an MCP server runs a tool call as a task, Claude Code now immediately moves it to the background and sends the result to Claude when it completes.

    Usage details

    Who can use it: Claude Code 2.1.319; terminal, gateway, and telemetry-disabled modes are supported.

    How to use it: Update Claude Code to 2.1.319 and use task calls from your MCP server. For the previous blocking behavior, use `CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND=1`.

    Tip: For long MCP operations, design a brief status message that tracks the task result so the terminal is not blocked.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • CapabilityClaude Code

    Sandbox escape during worktree creation fixed

    Fixed an issue that could bypass sandbox protections during worktree creation; on self-hosted runners, `--trust-workspace false` also disables repository settings and skill hooks.

    Usage details

    Who can use it: Claude Code 2.1.319; projects using self-hosted runners and sandboxes.

    How to use it: Update Claude Code; use `--trust-workspace false` with untrusted repositories.

    Tip: Inspect hook and skill files before running untrusted repositories.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingSelf-hosted runner

    Repository checkout directories no longer conflict in the same session

    `CLAUDE_RUNNER_CHECKOUT_PATH` now uses a separate directory for each repository in the same session; it appends suffixes such as `-2` and `-3` to conflicting names.

    Usage details

    Who can use it: Claude Code 2.1.319 self-hosted runner.

    How to use it: Continue using the environment variable; directories are separated automatically when multiple repositories are opened in the same session.

    Tip: In cleanup scripts, account for suffixed names as well as the fixed directory name.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • Pricing / PlanClaude apps gateway

    Priority rules added for multiple pooled budgets

    Gateway administrators can use `admin.pooled_conflict_rule` and `admin.pooled_priority_order` to determine which pooled budget is used first when a developer belongs to multiple groups.

    Usage details

    Who can use it: Claude apps gateway 2.1.319.

    How to use it: Add the `admin.pooled_conflict_rule` and `admin.pooled_priority_order` fields to the Gateway configuration.

    Tip: To manage budget transitions without surprises, test the rule with a multi-group test account before publishing it.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • APIClaude Code

    A latency-optimize command is now available for the Claude API

    The new `/claude-api latency-optimize` command finds bottlenecks in your API code that keep users waiting and tests faster settings against your eval set.

    Usage details

    Who can use it: Claude Code 2.1.319 and projects using the Claude API.

    How to use it: In a Claude Code session, enter `/claude-api latency-optimize` and run the recommended settings against your own eval set.

    Tip: Before deploying the recommendation, compare your quality metrics and time to first token together.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • RetiredClaude in Chrome

    Claude in Chrome shortcut listing tools removed

    `shortcuts_list` and `shortcuts_execute` have been removed from Claude Code’s built-in Claude in Chrome tools; they no longer list or launch shortcuts saved in the extension.

    Usage details

    Who can use it: Claude in Chrome integration with Claude Code 2.1.319.

    How to use it: Remove skill or plugin calls that depend on these tools and use the available browser tools.

    Tip: If you have your own automation, scan for tool name dependencies before upgrading.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • Pricing / PlanClaude Enterprise

    Spend tracking by cost code beta launches for Enterprise projects

    Eligible Enterprise organizations can tag projects with labels such as a cost code or customer file number and view spending by those tags.

    Usage details

    Who can use it: Claude Enterprise beta; unavailable with HIPAA, CMEK, ZDR, non-US data residency, and some custom provider configurations.

    How to use it: Create a tag at Organization settings > Organization > Project tags; assign a value in the project using Tags > Add tag. For Claude Code, add `attributionProject` to the settings.

    Tip: Add the tag about two hours before work begins; previous spending is not tagged retroactively.

  • Plugin / IntegrationClaude Code

    Servers using the legacy MCP handshake are fixed

    An issue that prevented servers using an older MCP handshake from connecting or returning empty results has been fixed; Claude Code no longer advertises to older servers an extension it uses only with newer servers.

    Usage details

    Who can use it: Claude Code 2.1.319; local and remote MCP servers.

    How to use it: Update Claude Code and restart your MCP server; no custom client setting is required.

    Tip: Still plan to migrate your server to the current protocol; legacy compatibility is not a long-term guarantee.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • Pricing / PlanClaude Code

    Periodic spending views now available for gateway users

    Claude apps gateway developers without a spending limit can see today’s, this week’s, and this month’s spending in `/usage`.

    Usage details

    Who can use it: Environments where the gateway runs Claude Code 2.1.319 or later.

    How to use it: Open `/usage` in Claude Code; make sure the gateway administrator is also running the same or a later version.

    Tip: Monitor periodic usage by users without limits instead of waiting for the end-of-month bill.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • APIClaude apps gateway

    Gateway operational logs can be output as JSON

    With `CLAUDE_GATEWAY_LOG_FORMAT=json`, gateway logs can be converted into JSON lines with `severity`, `time`, and `message` fields.

    Usage details

    Who can use it: Claude apps gateway 2.1.319.

    How to use it: Add `CLAUDE_GATEWAY_LOG_FORMAT=json` to the gateway process environment and enable field-based level parsing in your log collector.

    Tip: Update your staging log schema first; don’t leave legacy text matching in production.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingClaude Code

    A setting to reject calls instead of prompting for permission has been added

    The `permissions.denyInsteadOfPrompting` setting rejects tool calls that would otherwise wait for a permission prompt, allowing the turn to continue.

    Usage details

    Who can use it: Terminal sessions, subagents, and teammate sessions.

    How to use it: Add `permissions.denyInsteadOfPrompting: true` to your settings file.

    Tip: Use it to prevent waits for approval in unattended CI or long-running agent tasks; test deny rules before enabling broad coverage.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • AgentClaude Code

    Self-paced loop ends if it is not rescheduled

    Self-paced `/loop` no longer schedules a fallback wake-up when a turn ends without rescheduling.

    Usage details

    Who can use it: Claude Code 2.1.319.

    How to use it: If you need the previous behavior, set `CLAUDE_CODE_LOOP_KEEPALIVE=1`; with the new behavior, explicitly have `/loop` reschedule itself.

    Tip: Prefer the new default to prevent accidental infinite loops in persistent tasks.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • AgentClaude Code

    Completed subagents can be reopened from the Tasks screen

    The Earlier in this session row added to `/tasks` shows subagents that disappeared from the list or completed after a restart, and lets you message them again.

    Usage details

    Who can use it: Claude Code 2.1.319; terminal and relevant session views.

    How to use it: Open `/tasks`, select a subagent in the Earlier in this session row, and reopen it.

    Tip: For long tasks, check the main session’s latest context before reopening a subagent’s completed report.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • CapabilityClaude Code for VS Code

    Prompts over 50,000 characters are no longer silently truncated in VS Code

    Prompts exceeding 50,000 characters are no longer silently truncated; a warning remains in the chat box, and pastes over the limit are rejected with a clear notification.

    Usage details

    Who can use it: Claude Code VS Code extension 2.1.319.

    How to use it: Update the extension; split a long prompt into parts or attach it as a file.

    Tip: For code or data, specifying a file path and target section is safer than trying to fit it into a prompt.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingClaude Tag

    Models admin row added for Claude Tag

    For Enterprise organizations connected to Claude Tag, a Claude Tag row and a link to where model defaults are managed have been added to the Models admin page.

    Usage details

    Who can use it: For Claude Enterprise organizations connected to Claude Tag.

    How to use it: Open the Models page in admin settings and select the link in the Claude Tag row.

    Tip: When changing the Slack default model, also verify other defaults at the channel and workspace levels.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingClaude Code

    A variable to disable feature flag downloads has been added

    `CLAUDE_CODE_DISABLE_FEATURE_FLAGS` stops feature flags from being fetched but does not disable telemetry; each flag uses its built-in default.

    Usage details

    Who can use it: Claude Code 2.1.319.

    How to use it: Set `CLAUDE_CODE_DISABLE_FEATURE_FLAGS=1` in the runtime environment.

    Tip: Use it to keep behavior consistent in isolated or network-restricted environments, and recheck defaults with version updates.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes
  • SettingClaude apps gateway

    Organization message can be added to the gateway login screen

    The optional `login_notice` setting displays an operator message on the browser verification page and the Claude Code login screen.

    Usage details

    Who can use it: Claude apps gateway 2.1.319.

    How to use it: Enter the message to display in the `login_notice` field in the gateway configuration, then restart the gateway.

    Tip: Add a support link or maintenance announcement to the message; don’t include secrets or personal data.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 release notes

OpenAI · ChatGPT

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

New in this edition · 17

  • Plugin / IntegrationCodex CLI

    Codex CLI shows API key and cloud provider options under alternative sign-in

    Codex CLI 0.167.0 groups API key and cloud provider options under Sign in another way. When gcp_setup_wizard is enabled, Google Cloud Vertex AI can be set up with Application Default Credentials or a credentials file.

    Usage details

    Who can use it: Codex CLI 0.167.0; Google Cloud Vertex AI setup is available to users with the relevant accounts and credentials.

    How to use it: Select Sign in another way on the sign-in screen; for Vertex AI, enable the gcp_setup_wizard feature and complete setup with Application Default Credentials or a credentials file.

    Tip: Don’t put the credentials file in your repository; provide it through your local machine or a secure secrets store.

  • AgentCodex CLI

    Codex CLI adds an agent command center and unread result indicators

    Codex CLI 0.167.0 adds an agent command center and unread result indicators in the session picker. The command center can also be opened without a daemon.

    Usage details

    Who can use it: CLI users running Codex CLI 0.167.0.

    How to use it: Update with npm install -g @openai/codex@0.167.0 and open the command center with codex agents --no-daemon.

    Tip: For multiple long-running agent tasks, use unread indicators to prioritize completed results.

  • APIOpenAI Node SDK

    Node SDK delivers Agents session errors as stream events

    openai-node v7.53.0 enables Agents API session errors to be delivered as events during streaming. This lets the client handle the error without waiting for the session to end.

    Usage details

    Who can use it: Developers using OpenAI Node SDK v7.53.0 and the Agents API.

    How to use it: Upgrade to v7.53.0 with npm install openai; listen for session error events while consuming the Agents session stream, and pass them to the user or retry flow.

    Tip: Handle error events separately from regular output events in the stream consumer, and limit repeated retries.

    ✓ Verified by official sourceSource: openai-node sürümleri, v7.53.0
  • CapabilityCodex CLI

    Codex CLI closes network approval prompts when requests finish

    Codex CLI 0.167.0 closes the associated approval prompt when a network request ends. Concurrent requests to the same endpoint are also approved individually.

    Usage details

    Who can use it: Codex CLI 0.167.0 and tasks that request network permission.

    How to use it: Update the CLI; run a task that requests network access and expect approval prompts to close when the requests are complete.

    Tip: For concurrent network tasks, keep endpoints as narrowly scoped as possible to make it easier to track the number of prompts and the permission scope.

  • SettingCodex CLI

    Codex CLI adds macOS and Windows system proxy settings

    Codex CLI 0.167.0 adds an experimental option to use macOS or Windows system proxy settings.

    Usage details

    Who can use it: macOS and Windows users running Codex CLI 0.167.0; the option is in experimental settings.

    How to use it: Open the /experimental menu in the CLI and enable the option to use macOS or Windows system proxy settings.

    Tip: On corporate networks, test proxy compatibility with Codex traffic and authentication before enabling it.

  • FeatureCodex CLI

    Codex CLI opens patch details in the transcript on the approval screen

    Codex CLI 0.167.0 lets you expand patch details directly in the transcript while an approval prompt is open.

    Usage details

    Who can use it: Codex CLI 0.167.0 and tasks that require patch approval.

    How to use it: When an approval prompt appears for a patch, expand the patch details section in the transcript, review the changes, and then approve or reject them.

    Tip: Check the details before approving, especially for changes across many files.

  • CapabilityCodex CLI

    Codex CLI improves automatic context compaction for failed outputs

    Codex CLI 0.167.0 includes the stored output of failed responses and missing token usage data in its automatic context compaction calculation.

    Usage details

    Who can use it: Codex CLI 0.167.0 and long tasks that use automatic context compaction.

    How to use it: Update the CLI; use automatic compaction on a long task. No additional setting is required.

    Tip: Keeping interim updates concise and meaningful during long tasks helps preserve the necessary context after compaction.

  • Plugin / IntegrationCodex CLI

    Codex CLI adds backend switching and draft preservation from the session picker

    Codex CLI 0.167.0 lets you switch between connected backends from the terminal session picker. Drafts are preserved when switching backends.

    Usage details

    Who can use it: CLI users running Codex CLI 0.167.0 and using multiple connected backends.

    How to use it: Update the CLI to 0.167.0; open the terminal session picker and switch the connected backend, continuing with the current draft preserved during the switch.

    Tip: Use the session picker instead of deleting your draft before switching providers, so you don’t have to rewrite an unfinished prompt.

  • CapabilityCodex CLI

    Codex CLI preserves saved chats when switching providers

    Saved chats remain visible when the model provider is changed in Codex CLI 0.167.0. The transcript of a chat that can’t be continued can be read and carried over to a new chat as an editable prompt.

    Usage details

    Who can use it: Codex CLI 0.167.0 users.

    How to use it: Update the CLI; after switching providers, open the saved chat from the session picker. If it can’t be continued, read its transcript and start a new chat with an editable prompt.

    Tip: Check the chat transcript before switching providers to avoid losing important context.

  • CapabilityCodex CLI

    Codex CLI saves audio to history in long Realtime V3 conversations

    Codex CLI 0.167.0 saves completed speech segments to chat history as the transcript grows during long Realtime V3 conversations.

    Usage details

    Who can use it: Long conversations using Codex CLI 0.167.0 and Realtime V3.

    How to use it: Update the CLI and continue a Realtime V3 conversation; completed speech segments are added to the session transcript as it progresses.

    Tip: In long voice sessions, mark important decisions separately as text to make the searchable history more useful.

    ✓ Verified by official sourceModels: Realtime V3Source: ChatGPT ve Codex changelog, Codex CLI 0.167.0
  • CapabilityCodex CLI

    Codex CLI previews local image attachments as you compose

    Codex CLI 0.167.0 enables local image attachments to be previewed as you compose in supported terminals. Displayed local images can be reopened through clickable paths in the transcript.

    Usage details

    Who can use it: Codex CLI 0.167.0 and terminal environments that support image previews.

    How to use it: Update the CLI, attach a local image in a supported terminal, and check the preview before sending; reopen the image by clicking its path in the transcript.

    Tip: For code or design reviews, use the preview to verify that the image is the correct file before sending.

  • OtherCodex CLI

    Codex CLI fixes terminal clearing and draft editing behavior

    In Codex CLI 0.167.0, Ctrl+L clears the terminal view while previous messages remain accessible with Ctrl+T. When deleting text reveals a leading ! character, the draft no longer switches to command mode by mistake.

    Usage details

    Who can use it: Codex CLI 0.167.0 users.

    How to use it: Update the CLI; clear the terminal view with Ctrl+L and open history with Ctrl+T. The behavior when deleting a leading ! character in a chat draft is fixed automatically.

    Tip: In long sessions, use Ctrl+L only to declutter the view; previous messages are not deleted.

  • APIOpenAI Go SDK

    Go SDK recognizes GPT-6 Rosalind and image output controls

    openai-go v3.95.0 added GPT-6 Rosalind to the model catalog. The same release added image complexity and GIF and SVG output controls for image requests.

    Usage details

    Who can use it: Go developers using openai-go v3.95.0 with access to the relevant API models.

    How to use it: Update the Go module to v3.95.0; use GPT-6 Rosalind and the complexity, GIF, or SVG options in the generated API types for image generation/editing.

    Tip: After upgrading the SDK, run go vet and build tests for existing image requests.

    ✓ Verified by official sourceModels: GPT-6 Rosalind, GPT-Image-2.6 Sunburst, GPT-Image-2.6 FlareSource: openai-go sürümleri, v3.95.0
  • APIOpenAI Node SDK

    Node SDK recognizes GPT-6 Rosalind and image output controls

    openai-node v7.51.0 added GPT-6 Rosalind to the model catalog. The release also added image complexity and GIF and SVG output controls to the Node types.

    Usage details

    Who can use it: JavaScript/TypeScript developers using openai-node v7.51.0 with access to the relevant API models.

    How to use it: Install the latest SDK with npm install openai; use GPT-6 Rosalind and the image complexity, GIF, or SVG options in the Responses and image endpoints.

    Tip: Keep SDK types up to date in TypeScript projects so new fields are caught at compile time.

    ✓ Verified by official sourceModels: GPT-6 Rosalind, GPT-Image-2.6 Sunburst, GPT-Image-2.6 FlareSource: openai-node sürümleri, v7.51.0
  • APIOpenAI Python SDK

    Python SDK recognizes GPT-6 Rosalind and image output controls

    openai-python v3.46.0 added GPT-6 Rosalind to the model catalog. The same release added image complexity and GIF and SVG output controls for Responses image requests.

    Usage details

    Who can use it: Python developers using openai-python v3.46.0 with access to the relevant API models.

    How to use it: Upgrade to v3.46.0 with pip install -U openai; then use GPT-6 Rosalind and the complexity, GIF, or SVG options in the SDK’s Responses and image generation/editing types.

    Tip: Before deploying to production, validate the new fields with type checking and a small sample request.

    ✓ Verified by official sourceModels: GPT-6 Rosalind, GPT-Image-2.6 Sunburst, GPT-Image-2.6 FlareSource: openai-python sürümleri, v3.46.0
  • APIOpenAI Ruby SDK

    Model selection option added to Ruby SDK Responses requests

    openai-ruby v0.120.0 added model selection to the typed client for Responses requests.

    Usage details

    Who can use it: Ruby developers using openai-ruby v0.120.0.

    How to use it: Upgrade the gem dependency to v0.120.0 and explicitly select the model you want in the Responses call using the SDK’s new model field.

    Tip: Manage model selection through application settings to control model changes without deploying code.

    ✓ Verified by official sourceSource: openai-ruby sürümleri, v0.120.0
  • APIOpenAI Ruby SDK

    Ruby SDK adds data residency, GPT-6 Rosalind, image, and agent options

    openai-ruby v0.119.0 added data residency options for Australia, India, and Japan, and GPT-6 Rosalind to the model catalog. The release also expanded image complexity, GIF/SVG outputs, agent turn errors, and image/audio options.

    Usage details

    Who can use it: Ruby developers using openai-ruby v0.119.0; access to relevant regional residency options and API models depends on organizational permissions.

    How to use it: Upgrade the gem dependency to v0.119.0; use regional residency, GPT-6 Rosalind, image complexity/GIF/SVG, and agent image/audio options in the updated generated types.

    Tip: Test regional residency choices against the data governance requirements of the deployment environment.

    ✓ Verified by official sourceModels: GPT-6 Rosalind, GPT-Image-2.6 Sunburst, GPT-Image-2.6 FlareSource: openai-ruby sürümleri, v0.119.0

Meta · Meta AI

#8 Muse Wonder FullMuse Wonder
Full history (128) →

New in this edition · 3

  • FeatureMuse

    Personal workspaces created from ideas arrive in Muse

    Meta launched Spaces, a feature in Muse that turns an idea described by the user into a persistent, customized experience such as a tracker, planner, or project board. Spaces can be added to the sidebar or main tabs, hidden, and scheduled to appear at specific times; connected apps can be updated in the background.

    Usage details

    Who can use it: All locations where Muse is available; no additional country- or plan-based restrictions were specified.

    How to use it: Open Muse, describe the experience you want to create in natural language, and add the resulting Space to the sidebar or main tabs. If you like, connect apps, hide the Space, or schedule when it appears.

    Tip: Use a short description focused on a single purpose; reduce unnecessary data flows by connecting only the apps you want to keep up to date.

  • ModelMuse ve Meta AI

    Muse Wonder-based specialized models are rolling out for Muse and Meta AI

    Meta has begun rolling out specialized versions of existing model families trained using Muse Wonder to Muse and Meta AI users. These versions bring Active Memory and Active Focus capabilities; Meta reports that Muse forgets critical information 2.4 times less often and that performance on completing unexpected problematic tasks has increased by 67 percent.

    Usage details

    Who can use it: Muse and Meta AI users; the rollout may be gradual, and no country or plan distinctions were specified.

    How to use it: Use the usual chat and task flows on Muse or meta.ai; no new model selection or user setting is required, as the update is rolling out server-side.

    Tip: State long-term goals, persistent preferences, and task constraints explicitly; even though Active Memory carries context forward, verify critical decisions and outputs.

    ✓ Verified by official sourceModels: Muse Wonder, specialized versions of existing modelsSource: Meta AI Research: Bringing Specialized Models to Muse and Meta AISource: Meta Newsroom: Design the Muse App, Just for You
  • AvailabilityWhatsApp Calling API

    Quality-based business-initiated calling is automatically enabled for eligible numbers

    WhatsApp has begun automatically enrolling eligible business phone numbers in certain countries where the Calling API is enabled in the quality-based business-initiated calling model; existing eligible numbers are also being migrated automatically. This model allows business-initiated calls without prior approval, as well as gradually increasing call limits with persistent permissions, unless the user has refused persistent call permission or revoked permission they previously granted.

    Usage details

    Who can use it: Business phone numbers in certain countries that Meta deems eligible and have enabled the Calling API; the country list is not disclosed in the official announcement.

    How to use it: Enable the Calling API for your business phone number and subscribe to the calls webhook. Then check eligibility, user permissions, and limits in the Business-Initiated Calls flow; eligible numbers are automatically enrolled or migrated.

    Tip: After the automatic migration, monitor users' revocations of persistent permission, call quality, and gradually increasing limits; do not assume all numbers are eligible at the same time.

Moonshot AI · Kimi

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

New in this edition · 4

  • CapabilityKimi Code Web

    Fixed steer messages being incorrectly queued in active sessions

    Kimi Code 2.5.2 fixed an issue that incorrectly queued steer messages sent while a session was running. The steering message is now correctly delivered to the active Agent flow.

    Usage details

    Who can use it: Active sessions in Kimi Code Web; 2.5.2 and later.

    How to use it: Send a steer message in an active session in Kimi Code Web; on version 2.5.2 or later, confirm that the message is routed to the active Agent flow instead of waiting in the regular queue.

    Tip: Use the steer flow instead of a regular message to give new instructions during a long-running task; this lets you redirect the task without interrupting it.

  • CapabilityKimi Code Web

    Fixed project creation in Kimi Code Web over HTTP

    Kimi Code 2.5.2 fixed an issue that prevented users from creating a new project when accessing the Web interface over HTTP. This makes the project creation flow usable again in HTTP setups on a local network or behind a proxy.

    Usage details

    Who can use it: Users who access Kimi Code Web over HTTP instead of HTTPS; 2.5.2 and later.

    How to use it: Open Kimi Code Web using an HTTP address, start the workspace or project creation flow, and check that the new project is saved successfully.

    Tip: For internal installations that use HTTP or sit behind a proxy, specifically test creating a new project after upgrading.

  • Plugin / IntegrationKimi Code CLI

    Fixed duplicate tool errors caused by changes to MCP tool definitions

    Kimi Code CLI 2.5.2 fixed an error in which the select_tools flow reported the same tool twice when MCP tool definitions changed. This prevents sessions from being rejected during dynamic tool loading.

    Usage details

    Who can use it: Kimi Code CLI 2.5.2 and later; MCP servers and sessions using select_tools.

    How to use it: Upgrade Kimi Code CLI to version 2.5.2 or later; authorize your MCP server through /mcp-config and use tools through the select_tools flow.

    Tip: After updating the MCP server’s tool schema, start a new turn and check that the tool catalog refreshes without errors.

  • Plugin / IntegrationMicrosoft Foundry

    Kimi K3.1 Pro coming to the Microsoft Foundry model catalog

    According to Microsoft Foundry’s November 5, 2026 announcement, Kimi K3.1 Pro is being added to the Microsoft Foundry model catalog. The announcement lists text and image understanding, a 1 million token context window, and Foundry pricing (input $2.30, cache $0.23, output $11.50 / 1 million tokens).

    Usage details

    Who can use it: Developers using the Microsoft Foundry model catalog; for text-and-image workflows and a 1 million token context window with Kimi K3.1 Pro.

    How to use it: Open the Microsoft Foundry model catalog at https://ai.azure.com/catalog, select Kimi K3.1 Pro, and evaluate it first with a real coding, research, or document workflow.

    Tip: Establish a comparative baseline on your own workload for the quality of work completed with the current model, long inputs, visual content, and cost per task.

    ! UnverifiedModels: Kimi K3.1 ProSource: Microsoft Foundry Blog: Kimi K3.1 Pro

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 18

  • AvailabilityLoopForge

    LoopForge v0.3.0 releases a standalone Docker platform edition

    LoopForge v0.3.0 released a standalone platform edition with Main Bot workspace editing, persistent workspace checkpoints, Docker sandboxes, SOP execution, and human approval. The platform lets teams use CodeBuddy, Codex, or Claude Code to make code changes and deliver PRs/MRs through chat.

    Usage details

    Who can use it: Standalone Docker deployment; requires Git, curl, Docker Compose, and Python 3. The v0.3.0 GHCR image runs on Linux/amd64; GitHub integration is optional for changes to existing repositories.

    How to use it: Run `git clone --branch v0.3.0 https://github.com/Tencent/LoopForge.git && cd LoopForge` followed by `LOOPFORGE_IMAGE=ghcr.io/tencent/loopforge:0.3.0 LOOPFORGE_PUBLIC_URL=https://loopforge.example.com bash scripts/install.sh`; make the first account an administrator, then add a model key and executor.

    Tip: After installation, restrict IP access before exposing the service to the internet; LoopForge documentation limits the authentication boundaries of a new installation to trusted teams.

  • Capabilityteamai-cli

    teamai-cli moved all AI calls into workflow and scorer layers

    A breaking change routes AI calls for import, enrichment, maintenance, and session scoring through configurable default workflows; judgments are handled through `teamai-score` and configurable scorers. Direct launch APIs have been removed, and the `teamai/launch-through-workflows` lint rule has been added.

    Usage details

    Who can use it: Marked `Unreleased` on the official main branch; existing direct-launch integrations require migration. Default workflow files are under `workflows/`, and scorer definitions are under `scorers/`.

    How to use it: After updating the main branch, inspect workflows with `teamai list workflows`; configure custom behavior by editing the relevant `workflows/*.yaml` and `scorers/*.md` files, start a workflow with `teamai workflow start`, and run scorer operations with `teamai score` commands.

    Tip: For the breaking migration, first copy the default YAML and scorers and test them on a small workflow; remove direct AI launch calls before running lint.

  • CapabilityBrowserSkill

    BrowserSkill adds key rotation and recovery to local pairing

    A change published on the main branch adds mutual proof, persistent key rotation, and outage recovery to local pairing. The paired connection daemon verifies the connection before handling work traffic; device revocation and browser registration run in sequence, while pairing remains optional in connection settings.

    Usage details

    Who can use it: The change is listed as `Unreleased` on the official main branch; the CLI, extension, and DSH plugin can be released separately. Pairing is not required for new installations, and unverified connections continue to work with an identity warning.

    How to use it: Generate a connection with `bsk daemon pair` and paste it into the `Connection settings → Pairing link` field in the extension. Use `bsk daemon pair --replace` to replace it, `bsk daemon devices` to list devices, `bsk daemon revoke DEVICE_ID` to revoke a device, `bsk daemon rotate` to rotate keys, and `bsk daemon trust-local on|off` to manage local trust.

    Tip: To limit remote browser access, check the `bsk daemon devices` list regularly after pairing and revoke unused devices.

  • AvailabilityHy4

    Hy4 inference capacity reportedly increased tenfold

    TechNode reported that Tencent increased inference capacity tenfold on November 5, following heavy demand after Hy4's launch, and raised free usage quotas. No official Tencent confirmation page or updated quota table was found, so this item is verified only by secondary reporting.

    Usage details

    Who can use it: According to the report, WorkBuddy and CodeBuddy users are affected; the new quota amounts, regions, and plan coverage were not disclosed.

    How to use it: Select `Hy4` in the WorkBuddy or CodeBuddy model picker and retry the request; if the quota is still full, wait for a new official quota announcement and try again.

    Tip: Do not base production traffic solely on this secondary capacity report; roll out gradually while monitoring actual quotas and error rates in the application.

  • AvailabilityLoopForge

    LoopForge v0.3.1 switches to a pinned GHCR image instead of a local build

    With v0.3.1, the standard `./start.sh` flow pulls the versioned `ghcr.io/tencent/loopforge:0.3.1` image specified in `compose.yaml` instead of building locally. Separate `./start.sh dev` and `./start.sh release` flows are documented for source development and returning to the published image, respectively.

    Usage details

    Who can use it: LoopForge v0.3.1; the published image is for Linux/amd64. Use the source-build flow for other architectures.

    How to use it: Run `git clone --branch v0.3.1 https://github.com/Tencent/LoopForge.git && cd LoopForge && ./start.sh`; use `./start.sh dev` for source development or `./start.sh release` to switch back to the image.

    Tip: In production, use an immutable image digest instead of a tag; separate development and production stacks with the `dev` and `release` commands.

  • AvailabilityLoopForge

    LoopForge v0.3.2 changes to install Claude Code on first use

    In v0.3.2, Claude Code is installed on first use instead of being prebundled in the platform image. The same release also simplifies installation listening behavior by removing the `LOOPFORGE_PUBLIC_URL` requirement.

    Usage details

    Who can use it: LoopForge v0.3.2 Docker image; for platform installations that want to use Claude Code. The image is pinned as `ghcr.io/tencent/loopforge:0.3.2`.

    How to use it: Install with `git clone --branch v0.3.2 https://github.com/Tencent/LoopForge.git && cd LoopForge && ./start.sh`; wait for installation to finish the first time you select Claude Code as the executor in the interface.

    Tip: Before running production work, make the first Claude Code run with an empty SOP to verify the image's network and credential access.

    ✓ Verified by official sourceSource: LoopForge v0.3.2 resmi sürümü
  • Plugin / IntegrationLoopForge

    LoopForge v0.3.3 adds Codex and Claude Code SOP roles in Docker sandboxes

    v0.3.3 enables Codex and Claude Code SOP roles to run in Docker sandboxes. The release also fixes labeling for installation addresses and copying GitHub device codes over HTTP.

    Usage details

    Who can use it: LoopForge v0.3.3 and `ghcr.io/tencent/loopforge:0.3.3`; for installations using Docker sandboxes. Role executors can be set to Codex or Claude Code.

    How to use it: Update with `git clone --branch v0.3.3 https://github.com/Tencent/LoopForge.git && cd LoopForge && ./start.sh`; in SOP design, select Codex or Claude Code as the Agent role's executor and run it in a Docker sandbox.

    Tip: Verify separate model keys and sandbox permissions for each executor with a small test SOP.

    ✓ Verified by official sourceSource: LoopForge v0.3.3 resmi sürümü
  • Settingpage-harness

    page-harness v0.1.14 requires a per-run secret to resolve a review note

    v0.1.14 requires a per-run secret before a review note can be marked as resolved. The release notes also state that review text is labeled as untrusted data.

    Usage details

    Who can use it: page-harness v0.1.14, published on NPM; for installations using the review-note resolution flow. The change was verified against the official release page and commit.

    How to use it: Update to `page-harness@0.1.14`; start each review run with its own per-run secret value and send the same secret in the resolution request. Follow the existing integration's workflow for secret generation and transmission details.

    Tip: Do not put the secret in a URL or page text; transmit it only through the server-side authentication channel dedicated to that run.

  • AvailabilityTencent Cloud Enterprise Full-Stack Intelligent AI Suite

    Tencent Cloud Enterprise announces an eight-server Full-Stack Intelligent AI Suite

    In a November 6 PR Newswire release, Tencent Cloud Enterprise announced Full-Stack Intelligent AI Suite and Lite Deployment, combining TokenHub private MaaS, agent runtime sandbox/egress/credentials, and application layers. No official Tencent announcement or self-service activation page was found, so this item is verified only by a secondary distribution source.

    Usage details

    Who can use it: For enterprise and regulated-sector customers; Lite Deployment can start with eight servers, while an AI-only setup can start with three servers and is reported to run on existing hardware.

    How to use it: The announcement provides no self-service command or purchase URL; for a pilot, contact Tencent Cloud Enterprise sales or technical support to request an eight-server or three-server AI-only deployment.

    Tip: At the pilot stage, a three-server AI-only setup offers a lower threshold for measuring TokenHub and agent runtime requirements before a full eight-server installation.

  • Plugin / IntegrationWorkRally

    WorkRally CLI v2.12.0 added enterprise knowledge base search and writing

    WorkRally CLI v2.12.0 added commands to list and search enterprise knowledge bases, list cards, upload files, and write content after user approval. The feature is documented in SKILL.md for integration into Agent and CI workflows using an API key.

    Usage details

    Who can use it: WorkRally v2.12.0; knowledge base commands are for enterprise accounts only. An API key or OAuth session and `WORKRALLY_API_KEY`/configuration directory are required.

    How to use it: After running `npm install -g workrally` and `workrally auth login --token <YOUR_API_KEY>`, use `workrally kb list`, `workrally kb search --kb-id <id> --query "..."`, and `workrally kb upload ./dosya.pdf`; obtain user approval at every level before writing with `kb raw-create`.

    Tip: Keep search and writing as separate steps; show the target knowledge base and card to the user before writing to prevent data from being added to the wrong organization.

  • CapabilityWorkRally

    Draft mode enabled for supported models in WorkRally video generation

    WorkRally CLI v2.12.0 added a workflow that generates a low-resolution preview with video models that return `support_draft=true`, then upscales the final video using the same model at a higher resolution. The preview and final video are linked through `draft_id`.

    Usage details

    Who can use it: WorkRally v2.12.0; only video models with `support_draft=true` in the model catalog. The documented example uses `--resolution 1` for the draft and `--resolution 4` for the final generation with the same model.

    How to use it: First, use `workrally generate video-models` to select a model with `support_draft=true`; run `workrally generate video --prompt "..." --model <provider_id> --resolution 1 --draft '{"enable":true}' --poll`, then pass the output's `draft_id` to the same model with `--draft '{"draft_id":"<id>"}' --resolution 4`.

    Tip: Check the character, framing, and motion in the draft; start final-resolution generation only for an approved draft.

  • FeatureLoopForge

    LoopForge added shared AI schedules and role-based subagents

    The platform sync dated November 5 introduced shared AI scheduled-task templates that are versioned and bound to projects with parameters, and turned project roles within Main Bot into native subagents. It also added message queuing and steering, recovery of interrupted turns, GitHub diff/PR delivery status, and direct pushes to the shared branch.

    Usage details

    Who can use it: Released with the LoopForge v0.3.0 platform release; usable on your own Docker infrastructure with CodeBuddy, Codex, or Claude Code executors and an optional GitHub connection.

    How to use it: Install v0.3.0 and call a role with `@` in a Main Bot conversation; publish a scheduled task as a versioned template in the scheduled tasks section and bind it to a project and parameters. Queue a new message during a long response or steer it with `Steer`.

    Tip: Version recurring work as shared task templates; for critical deliveries, use a role subagent together with human approval.

  • Plugin / IntegrationAI-Infra-Guard

    AI-Infra-Guard v4.6.9 added a CloudGPU preset and broader MCP response compatibility

    v4.6.9 added a CloudGPU provider preset to Agent-Scan. MCP-Scan accepts XML attribute parameter tags from OpenAI-compatible models and mixed-case SARIF rule IDs; Agent-Scan can also handle null content deltas and SSE chunks containing only usage data.

    Usage details

    Who can use it: Listed as v4.6.9, dated 2026-11-05, in the official CHANGELOG; the GitHub Releases page still showed v4.6.8 at the time of checking, so verify the main branch/package status separately.

    How to use it: Install the latest A.I.G code, then use `pip install aig-agent-scan` and `aig-agent-scan --agent_provider target.yaml -o result.json` for Agent-Scan; for MCP validation, run `pip install aig-mcp-scan` and `aig-mcp-scan --repo /path/to/mcp-server -o result.sarif.json`.

    Tip: Verify the CloudGPU connection first with a low-scope `target.yaml`; in systems consuming MCP results, handle SARIF IDs in their now case-insensitively normalized form.

  • CapabilityAppCIF-Bench

    AppCIF-Bench v1.1 released an app-level instruction test with 466 examples

    Tencent-Hunyuan released the AppCIF-Bench v1.1 code, simulation engines, and Chinese-English datasets for real-world application tasks with long horizons and nested rules. The release contains 466 examples and 3.017 atomic rubric items; it also reports scores for 12 LLMs.

    Usage details

    Who can use it: Public GitHub repository; Python 3.10+, Apache-2.0 code license, and non-commercial CC BY-NC 4.0 data license. OpenAI-compatible endpoints or local model functions can be used.

    How to use it: After running `git clone https://github.com/Tencent-Hunyuan/AppCIF-Bench.git` and `pip install -r requirements.txt`, run `python run_eval.py pull_answer ...`, then `evaluate ...`, and finally `calculate_accuracy ...`.

    Tip: When comparing results, record dataset version v1.1, the judge model used, and the SSR and HSR metrics together.

    ✓ Verified by official sourceSource: AppCIF-Bench resmi GitHub deposu
  • CapabilityFlashTopK

    FlashTopK released as an exact, fast fused top-k kernel for sparse attention

    Tencent-Hunyuan open-sourced FlashTopK, which performs fast, exact, and deterministic fused top-k selection for learned sparse attention training and inference. The repository provides usage examples for supported indexer models, including Hy4-preview, and NVIDIA Hopper GPUs.

    Usage details

    Who can use it: Apache-2.0-licensed open-source repository; requires Python 3.10+, PyTorch 2.6+, Triton 3.3+, nvidia-cutlass-dsl 4.5+, and an NVIDIA H800, H20, H100, or H200.

    How to use it: Clone the repository and run `pip install -e .`; then use `flashtopk.select(...)` with `import flashtopk`, or choose the `flashtopk.Selector(...).plan(...).run(...)` workflow.

    Tip: Because the first call compiles the kernels, precompile with `Selector.plan()` before production workloads.

    ✓ Verified by official sourceModels: GLM-5.1, GLM-5.2, GLM-5.3, GLM-5.3-Flash, Hy4-previewSource: FlashTopK resmi GitHub deposu
  • Settingteamai-cli

    teamai benchmark arms gained an option for selective recall rollout

    In benchmark configuration, an arm can now take the team's recall rollout into its own trial clones with `recall: true`. The setting is not treated as a comparison matrix key; other workflow clones that do not specify it run without recall.

    Usage details

    Who can use it: Available on the official main branch; affects only benchmark arms set to `recall: true`, while unselected arms retain the default behavior.

    How to use it: Add `recall: true` to the relevant entry under `arms:` in the benchmark YAML; run the TeamAI benchmark with its usual command and do not add `recall` to the comparison keys.

    Tip: To measure the effect of recall, keep `recall: true` on both arms and change only the team content; otherwise, differences in results will conflate two interventions.

    ✓ Verified by official sourceSource: teamai-cli benchmark recall commit'i
  • ModelWeLM-Image

    WeLM-Image introduced as a unified image generation model with 31.7B parameters

    A November 5 paper by Tencent WeChat Vision describes WeLM-Image, a 31.7B-parameter model that combines text-to-image generation, instruction-based editing, and multi-reference workflows in a single model family, along with a faster WeLM-Image-Turbo variant. I could not find an official model repository, weights, or a usable API; the claim that the model is open is supported only by the paper and secondary coverage.

    Usage details

    Who can use it: Research paper only for now; no verified publicly available weights, demo, API, or regional/plan access.

    How to use it: No instructions for running it can be provided because no official usage path has been published; review only the arXiv paper and wait for an official weights/endpoint announcement.

    Tip: Do not treat the ‘open’ wording in secondary coverage as verified access to a downloadable model; check the official repository and license before including it in a production plan.

    ! UnverifiedModels: WeLM-Image, WeLM-Image-TurboSource: WeLM-Image arXiv makalesiSource: AiWeekly WeLM-Image haberi
  • SettingTokenHub-Vendor-Verifier

    TokenHub Vendor Verifier adds 16 concurrency points for the 200k test

    The performance test adds a concurrency point of 16 to the default concurrency ladder for the 200k-token bucket; the upper limits for long-context buckets are narrowed to 32 for 128k and 8 for 512k and 1m. If a custom `CONCURRENCY_LADDER` is provided, the user-defined ladder is preserved, and unknown buckets are skipped instead of silently falling back to 128.

    Usage details

    Who can use it: The `thvv/perf` tool on the main branch; any API compatible with OpenAI Chat Completions or Anthropic Messages. The default `TPM_BUDGET` is now 0, meaning rate limiting is disabled.

    How to use it: After running `pip install -r requirements.txt` under `thvv/perf` and configuring `.env`, run `bash run.sh bench-all`; for your own ladder, use `CONCURRENCY_LADDER="1 8" bash run.sh bench-all`.

    Tip: The 200k×16 point may exceed the real gateway TPM quota; treat 429 responses as capacity observations rather than errors, and use evalscope `--rate` for a hard rate limit.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 22

  • CapabilityQwen Code Managed Agent

    Managed Agent Files/Shell sessions get glob, todo_write, and PDF pages

    Ordinary Managed Files sessions now have native glob, Shell sessions have persistent todo_write, and Hosted read_file supports PDF pages ranges. Saved todo state is reloaded, and oversized PDF results now prompt a retry with a smaller page range.

    Usage details

    Who can use it: In Ordinary Managed Agent `files/1` and `shell/1` profiles and Hosted read_file; Plan mode retains its todo-writing restriction, while resident/CSI and SSH profiles retain their stated limitations.

    How to use it: Match files with glob in a Files session; use todo_write in a Shell session; provide a pages range in a read_file request when reading a PDF. Keep the page range to no more than 20 pages.

    Tip: Request only the page range you need instead of reading a large PDF in a single request; this reduces both context usage and retry costs.

  • Plugin / IntegrationQwen Code GitHub Channel

    PR-bound sessions receive plain comments from allowlisted senders without a mention

    When deliverToBoundSessions is enabled, a plain conversation comment on a PR from an account on the sender allowlist can now be delivered to the bound session monitoring the PR without an @mention. This special path does not generate a public reply; the groups.*.requireMention rule remains in effect for public replies.

    Usage details

    Who can use it: In configurations that monitor PR-bound sessions in GitHub Channel and use a repository group sender allowlist; the target rules for first contact, assignment, and review-request flows remain unchanged.

    How to use it: Enable deliverToBoundSessions in the GitHub Channel configuration and add the permitted sender account to the allowlist for the relevant repository group; comments on the bound PR session are routed through the private delivery path.

    Tip: Keep the public mention requirement enabled; this ensures plain comments from allowlisted senders go only to the bound session and do not trigger a public reply on PRs without a binding.

  • FeatureQwen Code

    Local session hooks can monitor FileChanged notifications

    In Local and non-managed sessions, FileChanged hooks report file changes in the working directory as add, change, or unlink. Absolute files or recursive directories can be registered through SessionStart and FileChanged output; changes made by external processes are also monitored.

    Usage details

    Who can use it: In Local, non-managed Qwen Code sessions; remote, sandboxed, or Managed runtimes do not start a host file watcher.

    How to use it: Register the paths to monitor as watchPaths in SessionStart hook output and define a FileChanged hook; read the file_path and event type in the notification to start your own process.

    Tip: Avoid setting up a process that writes back to a directory monitored by the FileChanged hook, or filter its writes; otherwise, self-triggering loops can occur.

  • FeatureQwen Code Web Shell

    Session search moves to the Web Shell top row next to View activity

    Session search is no longer a control visible only on hover in the Workspaces heading; it is now always visible on the sidebar top row, immediately to the right of the View activity bell. The bell stays fixed when the search field opens, the field spans the full row on narrow screens, and the control is hidden in the activity view.

    Usage details

    Who can use it: In the Qwen Code Web Shell sidebar; the Channels section and narrow drawer/column layouts are also adapted.

    How to use it: Click the magnifying glass icon on the sidebar top row and enter a session name in the search field; press Escape or click the icon again to close it.

    Tip: When searching across many workspaces, first make sure the View activity view is closed; the search control is hidden while the activity panel is open.

  • FeatureQwen Code Web Shell

    Skill completion details show the name and Extension information separately

    In Web Shell Skill suggestions, long qualified headings now wrap to show the full Skill name, and the owning Extension moves to a separate line. Standalone Skill names and names containing colons are preserved; selecting a suggestion continues to insert the fully qualified command.

    Usage details

    Who can use it: In Daemon-backed Web Shell sessions and the workspace Skill catalog; in the nightly release.

    How to use it: Open slash suggestions in Composer and hover over an Extension Skill; read the Skill name and the Extension line below it in the details panel, then select the suggestion.

    Tip: For long qualified commands, first check the Extension owner on the separate line, then select the full qualified command to preserve the current namespace.

  • CapabilityQwen Code

    A changed prefix is revalidated as the local Managed journal grows

    If the local journal file grows within the same timestamps and changes a previously read prefix, Qwen Code now rereads that prefix to verify consistency. This reduces the risk of accepting an invalid transcript during concurrent writes and of showing an active session without a header.

    Usage details

    Who can use it: In Qwen Code sessions using local Managed journals and transcripts; the update is in the nightly release.

    How to use it: Upgrade to the Qwen Code nightly release; when another process writes to the journal, session list or writer lease reads will automatically use the new consistency check.

    Tip: For automations that modify journal files externally, keep append and rewrite operations under a single writer where possible; the new check catches races but does not replace workflow design.

  • APIQwen Code

    Large Skill ZIPs can be installed through the daemon API, and the result can be queried

    The daemon can install a Skill ZIP archive directly as binary content or by using an absolute file path on the daemon host. If the connection is lost, query the installation receipt using the attempt ID to determine whether the result was committed, not_committed, or recovery_required.

    Usage details

    Who can use it: For the global User scope or a selected workspace on daemons that report native archive capability; the compressed archive limit is 50 MiB, extracted content is 200 MiB, ZIP entries are 4,096, and the root Skill document is 32 MiB.

    How to use it: Check the capabilities response for native Skill archive support. Send the ZIP as application/octet-stream or provide an absolute path on the daemon host and, optionally, a SHA-256. Save the generated attempt ID and, if the response is lost, query the receipt using the same ownership context.

    Tip: If the response is lost, check the receipt before uploading again; repeating the same attempt ID does not reinstall it.

  • SettingQwen Code Managed Agent

    Managed Agent automation is enabled by default when no setting is specified

    Managed Agent automation is now enabled if the deployment doesn’t specify the automation flag; persistent definitions can be scheduled and run manually. An explicit false value preserves the existing behavior by disabling the browser and mutation routes.

    Usage details

    Who can use it: In Managed Agent deployments configured with Session Store; opt out with QWEN_MANAGED_AGENT_AUTOMATION_ENABLED=false or qwen.managed-agent.automation.enabled=false.

    How to use it: To use automation, start a deployment with Session Store enabled and no setting specified. To keep it disabled, set the environment variable or Spring property to false and restart the deployment.

    Tip: If you don’t want scheduled tasks to start before upgrading, explicitly set a false override; the default is now opt-out rather than opt-in.

  • AgentQwen Code Managed Agent

    Persistent structured user questions are now available in Managed Agent sessions

    In ordinary Managed Agent Files and Shell root sessions, the model can now start an ask_user_question flow with 1–4 questions and single-choice, multiple-choice, or custom-text answers. Unanswered questions remain pending in Full Access or Plan mode, are restored after a page refresh or Harness restart, and the original tool call can continue once answered.

    Usage details

    Who can use it: In ordinary Managed Agent Files/Shell root sessions, including Public Plan mode; the initial scope excludes MCP, Hook, sub-sessions, resident CSI, and foreground-only profiles.

    How to use it: In a Managed Agent Web Shell session, have the model ask for a decision or choice; submit a single-choice, multiple-choice, or custom-text answer in the form that appears. You can reopen and complete a pending form in the same session after refreshing the page.

    Tip: Use structured questions with option IDs instead of free text for branches where the model needs to reduce uncertainty; this lowers the risk of repeating the wrong choice after a restart.

  • FeatureQwen Code Web Shell

    The Web Shell Skills page adds large ZIP uploads and a way to check uncertain results

    The Skills page can now upload actual Skill ZIPs within the limits declared by the daemon, and reports success only after installation is confirmed. If the connection drops, it queries the same attempt with Check result instead of automatically resending the archive; activation errors are shown separately from installation success.

    Usage details

    Who can use it: In Web Shell sessions using a daemon that supports native Skill archive; the legacy base64 path remains available for small ZIPs on older daemons.

    How to use it: Open the Skills page in Web Shell, select the ZIP, and upload it. If the result remains uncertain, click Check result. If the status cannot be confirmed, use Stop checking to end local polling and verify the server-side result before uploading again.

    Tip: If the installation response is lost, don’t immediately upload the same ZIP again; first use Check result to confirm whether it may have been committed.

  • Plugin / IntegrationQwen Code Channels

    The Channel timeline distinguishes private delivery and scan health statuses

    The Channel event timeline now shows separate terminal records for events privately delivered to a PR-bound session, events consumed without delivery or retired because their target was lost, and degraded states where scanning has stopped. Each bound event row also identifies its target session.

    Usage details

    Who can use it: For Qwen Code Web Shell users monitoring PR-bound sessions with GitHub Channel.

    How to use it: Open the Channel event timeline view in Web Shell; inspect the target Session on private delivery rows, the reason for undelivered events, and the binding scan health rows.

    Tip: Check the terminal outcome and target Session before treating a `running` status as delivered; distinguish private-path records from degraded scan records in particular.

  • SettingQwen Code daemon

    Daemon event replay memory can be capped with --event-ring-max-bytes

    The daemon command and Skill replay ring can now also be capped by a byte budget; the default limit is 8 MiB, and the maximum is 256 MiB. Repeated delivery of identical command snapshots is reduced, and the latest complete command snapshot is preserved for reconnecting clients.

    Usage details

    Who can use it: In the Qwen Code serve daemon, fast-path, and related bridge startup paths; the default is 8 MiB, and the maximum allowed value is 256 MiB.

    How to use it: Pass the `--event-ring-max-bytes <bayt>` flag when starting the daemon; if no value is specified, the default of 8 MiB applies.

    Tip: Use a lower byte budget for long sessions under memory pressure; tune the limit based on measurements to avoid losing snapshots in heavy replay and reconnect scenarios.

  • AvailabilityQwen Code

    Image input for additional open vision model families arrives in stable v0.26.9

    Qwen Code model modality mapping now recognizes image input for Llama 3.2 Vision, Llama 4, LLaVA, Pixtral, and Mistral Small 3.1/3.2. Text-only versions of the same families aren’t automatically treated as vision models; video and PDF support are not part of this change.

    Usage details

    Who can use it: In Qwen Code sessions using these models under their Ollama, Hugging Face, or OpenAI-compatible server names.

    How to use it: Select the correct name for the vision model and attach an image to the message; Qwen Code now automatically declares the image modality for these models.

    Tip: Choose an explicitly named Vision, LLaVA, Pixtral, or Mistral Small 3.1/3.2 variant instead of a text-only tag such as `llama3.2:3b`.

    ✓ Verified by official sourceModels: Llama 3.2 Vision, Llama 4, LLaVA, Pixtral, Mistral Small 3.1, Mistral Small 3.2Source: Qwen Code v0.26.9 sürüm notlarıSource: Qwen Code açık görsel model modality PR’ı
  • CapabilityQwen Code Managed Agent

    Lock overhead reduced for Managed Agent refreshes, history reads, and tool execution

    v0.26.9 speeds up refreshes by indexing the activation lease lookup and speeds up history reconstruction by caching Managed Session message bodies per operation. SSH recording and worktree setup are less serialized; tool execution admission/dispatch locks, native journal snapshot locks, and the writer recovery path are also narrowed; MCP Google auth providers are loaded on demand.

    Usage details

    Who can use it: In Qwen Code deployments using Managed Agent hosted sessions, SSH workspace setup, MCP Google auth, and the daemon/Runtime; included with the release and requires no user setting.

    How to use it: Upgrade to Qwen Code v0.26.9; the changes apply automatically during activation refresh, history loading, SSH workspace setup, MCP startup, and tool execution.

    Tip: After upgrading, compare startup and first-response times, especially for Managed Sessions with long histories and SSH tasks, to measure the improvement.

    ✓ Verified by official sourceSource: Qwen Code v0.26.9 sürüm notları
  • CapabilityQwen Code

    Qwen VL models using Ollama names now accept images and video

    Ollama spellings of Qwen3-VL and Qwen2.5-VL models now match the image and video modality, just like their Hugging Face and DashScope names. In particular, `qwen3-vl:8b`, `qwen2.5vl:7b`, and the size-unspecified `qwen2.5-vl` names no longer fall back to the text-only path.

    Usage details

    Who can use it: For Qwen VL models connected to Qwen Code through Ollama.

    How to use it: Select one of these model names in the Ollama provider and attach an image or video to the message; no additional model name rewriting is required.

    Tip: Don’t try to fix the missing hyphen in the Ollama tag spelling `qwen2.5vl`; Qwen Code now recognizes this actual model name directly.

    ✓ Verified by official sourceModels: qwen3-vl:8b, qwen2.5vl:7b, qwen2.5-vlSource: Qwen Code v0.26.9 sürüm notlarıSource: Qwen Code Ollama Qwen VL modality PR’ı
  • APIQwen Code

    Skills can be activated in bulk with a single request for 1–100 names

    The Qwen Code daemon now includes a PUT /skills/activation endpoint that atomically enables or disables User-scoped Skill preferences. The TypeScript SDK’s setUserSkillsActivation also returns the saved preferences, per-name results, and propagation status.

    Usage details

    Who can use it: For User-scoped calls using Qwen Code daemon/serve; a single request processes 1–100 explicitly named Skills, and changes propagate to existing sessions.

    How to use it: Send a PUT /skills/activation request to the daemon or use setUserSkillsActivation in the TypeScript SDK. In the request body, provide the names of the Skills to enable or disable, then check the propagation status in the response.

    Tip: Send only the exact names you want to change in the same request; unrelated Skill restrictions and other settings are preserved.

  • APIQwen Code TypeScript SDK

    The TypeScript SDK supports Skill ZIP installation and installation receipt queries

    The public TypeScript SDK can now install Skill ZIPs from a File, Blob, byte array, or file path on the daemon host. By preserving the attempt ID in connection or timeout errors, the SDK lets you determine later whether the installation was actually processed.

    Usage details

    Who can use it: For global User Skills and a selected workspace; a daemon that supports native archive capability is required.

    How to use it: Use the native archive options of the installWorkspaceConfigSkill call in the SDK; provide File, Blob, or byte content, or a host path. If an error occurs, save the returned attempt ID and verify the result using the SDK’s installation receipt query.

    Tip: For ZIPs around 20 MiB, send them directly as a File or Blob instead of converting them to base64 JSON to reduce request size and memory usage.

  • CapabilityQwen Code

    Sessions with the same modification time are no longer skipped in the list

    When multiple session files have the same modification time, listSessions pagination no longer skips some sessions. This fixes sessions—especially those created quickly or updated simultaneously—not appearing in Web Shell history.

    Usage details

    Who can use it: Persistent Qwen Code session lists and the Web Shell flows that use them.

    How to use it: Refresh the session list or go to the next page; sessions with matching timestamps will appear as part of normal pagination.

    Tip: If a newly opened session isn’t visible, check the next page before refreshing; the new version should no longer skip entries.

    ✓ Verified by official sourceSource: Qwen Code v0.26.9 sürüm notları
  • CapabilityQwen Code Web Shell

    Web Shell splits multi-session deletion by route limits

    When deleting multiple Web Shell sessions, requests are batched according to daemon route limits instead of being compressed into a single oversized call. This makes deletion more reliable for large session lists.

    Usage details

    Who can use it: For users deleting multiple sessions from the Web Shell session list.

    How to use it: Select multiple sessions in the session list and run the delete action; Web Shell splits the deletions into groups that fit the route limit in the background.

    Tip: Instead of deleting a very large list all at once, first separate out active sessions; this reduces the risk of deleting the wrong session.

    ✓ Verified by official sourceSource: Qwen Code v0.26.9 sürüm notları
  • FeatureQwen Code Web Shell

    The Web Shell composer Add menu is clearer and more accessible

    The icons, labels, descriptions, spacing, and accessibility presentation in the composer’s Add menu have been refreshed. The Plan status indicator is also displayed more consistently in the menu.

    Usage details

    Who can use it: In sessions using the Qwen Code Web Shell composer.

    How to use it: Open the Add menu in the Web Shell message composer; actions are listed with the new icons, descriptions, and Plan status indicator.

    Tip: Check the indicator in the Add menu before sending to avoid switching to the wrong operating mode.

    ✓ Verified by official sourceSource: Qwen Code v0.26.9 sürüm notları
  • CapabilityQwen Code Web Shell

    Web Shell preserves unsent drafts while navigating history with the arrow keys

    When navigating wrapped prompt lines with ArrowUp or ArrowDown, the current unsent draft is no longer lost. This prevents line navigation from overwriting text you’re writing while reviewing messages from previous sessions.

    Usage details

    Who can use it: In all sessions using wrapped or multiline drafts in the Web Shell composer.

    How to use it: Type text in the composer, make it long enough to wrap across lines, and navigate history with ArrowUp/ArrowDown; return to the draft and continue editing it.

    Tip: It’s safe to navigate history before sending long commands, but backing up critical text in a separate editor is still a good habit.

    ✓ Verified by official sourceSource: Qwen Code v0.26.9 sürüm notları
  • AgentAI Gateway

    Agent management and session-level FinOps observability are now available in AI Gateway

    Alibaba Cloud AI Gateway Serverless now manages Agent Access and Model Access configurations in a single Agent resource. Usage, token, credit, tool call, and latency data can be inspected down to the Session, Turn, and model call levels; Agent statistics and intelligent insights are also available.

    Usage details

    Who can use it: Only in the Standard and Enterprise editions of the new Serverless gateway; Claude Code, Codex, DeepSeek Harness, Dify, Model Studio, and Custom Agent types are supported. Dedicated Edition and older Serverless gateways are not included.

    How to use it: In the AI Gateway console, open Agent Management > Create Agent for the target instance. Configure Agent Access or Model Access in Agent Topology. For session details, enable session analysis under FinOps > Session Usage Statistics; for Agent statistics, enable daily analysis under Agent > Statistics.

    Tip: First set up the consumer-to-Agent binding correctly so that token, error, and latency usage for Agents using the same Model API doesn’t get mixed together.

DeepSeek · DeepSeek

DeepSeek V4.1 Flash 1017
Full history (251) →

New in this edition · 1

  • AvailabilityDatabricks Foundation Model APIs

    DeepSeek V4 Flash (0731) access discontinued on Databricks

    Databricks retired DeepSeek V4 Flash (0731) from its Foundation Model APIs pay-per-token service as of November 5, 2026, and designated DeepSeek V4.1 Flash as the recommended migration model. Since there is no separate announcement about this date on DeepSeek's official channels, this entry confirms only the new availability status on the Databricks platform.

    Usage details

    Who can use it: Workloads using Databricks Foundation Model APIs pay-per-token; DeepSeek V4 Flash (0731) is unavailable after November 5, 2026, and the recommended alternative is DeepSeek V4.1 Flash.

    How to use it: Replace the model with DeepSeek V4.1 Flash in your Databricks model selection or serving configuration. To find workloads using the old model, adapt the query on Databricks' model retirement policy page to search for DeepSeek V4 Flash (0731).

    Tip: If you need long-term version pinning, consider Databricks' recommended provisioned throughput option and test the migration before the retirement date.

    ! UnverifiedModels: DeepSeek V4 Flash (0731), DeepSeek V4.1 Flash (1017)Source: November 2026 | Databricks on Google CloudSource: AI models maintenance policy | Databricks

Watchlist

Muse Wonder

44
Meta · Meta AI

Meta's standard Muse Wonder model; ranked 11th with 44 points, 2 points short of the TOP 10.

Claude Haiku 5.5

36
Anthropic · Claude

Released October 7; a fast, low-cost option at $0.10 / $0.50.

Grok 4.7

35
xAI · Grok

The default model in the Grok app. The default watermark for text models supported by the Grok API goes live today (November 7); there were no new changes during this period.

Tencent Hy4

35
Tencent · Hy

Available via API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, free on WorkBuddy and CodeBuddy through November 15. According to TechNode, inference capacity was increased tenfold (not officially confirmed).

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba's cloud flagship model; Qwen 4 has not been released yet. Qwen Code v0.26.9 adds Skill ZIP installation, structured questions, and Ollama Qwen VL support.

DeepSeek V4.1 Flash 1017

33
DeepSeek · DeepSeek

MIT-licensed open weights, 1M context, 33 points. Databricks retired the older V4 Flash (0731) on November 5 and recommended switching to V4.1 Flash.

Sources