YZ·RADARyz.mikail.net
Daily edition · 8 November 2026

Qwen Code 0.27 and the SLS agent layer; watermarking by default in the Grok API

No change in the TOP 10 for a fourth day: the rankings and scores of all ten models are unchanged from 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 LMArena text ranking is still dated October 27, and Gemini 4 Carbon is also in first place there. The watchlist is unchanged too: Muse Wonder is 2 points outside the TOP 10 with 44 points. This report’s scan was conducted on November 8 between 08:05 and 08:11, and 79 new developments were added.

Most of the activity comes from two developer tools. On the Anthropic side, the items missing from yesterday’s Claude Code 2.1.319 release notes have been filled in: the gateway now warns at startup about conflicting model IDs, mixing GovCloud with other upstreams, and base URL environment variables; the spending-limit error shows the reset date; the self-hosted runner no longer automatically uploads the debug log for a failed session, so the log stays on the host; and deny/ask rules written using connector names in Claude Desktop are now honored. On the Qwen side, Qwen Code v0.27.0 is out: Managed Agent sessions can now queue multiple sub-questions at once, GitHub Channel notifications can be filtered by repository using an allowlist/blocklist, Web Shell displays a 44-week activity heatmap for long-running sessions, and the interactive CLI reports status to the terminal via OSC 7501. Alibaba Cloud also announced an agent-era package for SLS: SessionStore, semantic layer, DataAgent, an OAuth/RAM-controlled MCP Server, and a Skills framework.

There are also smaller but practical changes. Google Cloud cut the price of Gemini Omni 1.2 Flash 720p video by 30% compared with Omni 1.1 Flash, to $0.07 per second. Starting November 7, xAI is adding watermarks by default to API text outputs from the grok-4.7, 4.6, 4.5, 4.3, and grok-4.20 models; the `x-watermark` header in the response indicates this, and team administrators can turn it off in Console. Meta paused updates to its commercial terms, catalog, and generative ads terms that were set to take effect on November 20. Moonshot has started removing legacy REST transcript endpoints and streaming conversations as turns/items in Kimi Code 2.6.0 canary builds; stable 2.5.2 is unaffected. At Tencent, teamai-cli v0.30.0-beta.0 added workflow ownership and `agentsmd/` support, while LoopForge added network environment selection for CodeBuddy. No new changes were found for OpenAI, DeepSeek, Z.ai, or Xiaomi during this period.

Today’s actions: If you use the Grok API, start recording the `x-watermark` header in your responses and decide whether watermarking works for you. If you have a client connecting to a Kimi Code server with a canary release, switch from transcript endpoints to the turns endpoint and the `subscribe_v2` subscription. Claude gateway administrators should restart the gateway after updating and read the startup warnings; Qwen Code users can add noisy repositories to the blocklist in GitHub Channel.

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 for a fourth day with 60 points. Access is still limited to selected cyber defenders and Fairwind partners; general API and app access have not yet opened.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. The remaining Claude Code 2.1.319 items clarify gateway, self-hosted runner, and Desktop connector behavior.Released: 22 September 2026

  3. 03

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

    OpenAI’s most powerful GPT-6 model, for long and difficult reasoning tasks. No new changes on the OpenAI side during this period.Released: 3 September 2026

  4. 04

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

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

  5. 05

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

    The most powerful model also available on the free claude.ai plan; 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 admin 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 became 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. Meta paused the commercial terms updates set to take effect on November 20; the current terms remain in force.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). In Kimi Code 2.6.0 canary, the server API switched from transcripts to a turns/items stream.Released: 23 October 2026

Key signals

  • The TOP 10 rankings and scores have remained unchanged for a fourth day: Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57. The watchlist is unchanged; 79 new developments were added today.
  • Anthropic: The remaining Claude Code 2.1.319 items have been added. The gateway warns about configuration conflicts at startup, the spending-limit message provides the reset date, runner debug logs now stay on the host, and Desktop honors connector rules.
  • Qwen: Qwen Code v0.27.0 brings concurrent sub-questions, a repository allowlist/blocklist filter for GitHub Channel, an activity heatmap in Web Shell, and status reporting via OSC 7501. Alibaba Cloud added SessionStore, semantic layer, DataAgent, MCP Server, and a Skills framework to SLS.
  • Google: 720p video generation with Gemini Omni 1.2 Flash is 30% cheaper, at $0.07 per second. xAI: Text outputs from supported models in the Grok API have been watermarked by default since November 7 (the `x-watermark` header).
  • Meta paused the commercial terms updates scheduled for November 20. REST transcript endpoints have been removed in Kimi Code 2.6.0 canary; conversations are streamed as turns/items (`subscribe_v2`).
  • Tencent: teamai-cli v0.30.0-beta.0 (workflow ownership and handoff, `agentsmd/`, Hermes transcripts); CodeBuddy network environment selection in LoopForge. No new changes at OpenAI, DeepSeek, Z.ai, or Xiaomi.

Worth trying today

01

Check watermarking in your Grok API outputs

Since November 7, API text outputs from the grok-4.7, grok-4.6, grok-4.5, grok-4.3, and grok-4.20 models have been watermarked without requiring code changes; structured outputs and tool-call arguments are also covered. Log the `x-watermark` header in your responses: `true` indicates that the output is marked. If your team does not want this, an administrator can uncheck the relevant model in Console → Team settings → Model output watermarking and save; the change takes effect for new requests in about 5 minutes.

xAI · Grok
02

Reduce noise in the Qwen Code GitHub channel

Update Qwen Code to v0.27.0. In Web Shell, go to Settings → Channels → GitHub, set Repository Filter to Allowlist or Blocklist, and add repositories in the Filtered Repositories field using the `owner/repo` format or a GitHub URL. Filtered notifications appear in the timeline as `repo_filter`. If you also use Group Policy, note that the active set of repositories is the intersection of the two.

Alibaba · Qwen
03

Create video drafts more cheaply with Omni 1.2 Flash

In Google Cloud Console, select the `gemini-omni-1.2-flash` model under Agent Platform Studio → Multimodal, or use this model name in your API request. Generate initial drafts at 720p ($0.07 per second), then upscale the best clip to 1080p or 4K. Check current rates on the Model Garden pricing page; allowlist approval is required for audio-referenced generation.

Google · Gemini
04

Review Claude gateway and runner settings

After updating Claude Code to 2.1.319, restart the gateway and read the startup logs: conflicts between custom model IDs, mixing GovCloud with other upstreams, and base URL environment variables now trigger warnings. Define the `admin.blocked_message` text so users can see why they hit a limit; the error now also shows the reset date. If you use a self-hosted runner, debug logs for failed sessions now stay on the host; collect them, and use `--remove-session-state` deliberately when removing them.

Anthropic · Claude

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 · 1

  • Pricing / PlanGemini Omni 1.2 Flash

    Gemini Omni 1.2 Flash 720p video price drops to $0.07 per second

    Google Cloud announced a 30% reduction in the price of 720p video generation with Gemini Omni 1.2 Flash compared with Gemini Omni 1.1 Flash, bringing it to $0.07 per second. Similar savings were reported for other resolutions; the model is ready for production use.

    Usage details

    Who can use it: Google Cloud API and production use; Provisioned Throughput is available. Generation with audio references requires allowlist approval.

    How to use it: Select the `gemini-omni-1.2-flash` model in Google Cloud Console > Agent Platform Studio > Multimodal, or use this model name in an API request; check current rates on the Google Cloud Model Garden pricing page.

    Tip: To reduce costs, generate initial drafts at 360p or 720p, upscale the best clip to 1080p or 4K, and reuse the same clip for edits with dialogue.

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 · 28

  • AgentClaude Code

    Agent reports, the Agents view, and cleanup flow are preserved

    Internal markers are hidden in the normal view of subagent reports; missing or rejected reports explain why and what to do next. The Agents view preserves pending permission responses, drafts, restart inputs, and long previews. /clear does not kill running background agents, teammates, or workflow commands; a safety stop recommends retrying with a narrower prompt.

    Usage details

    Who can use it: Claude Code sessions using subagents, teammates, the Agents view, and background workflows.

    How to use it: Review agent reports in normal/verbose view, track work with /tasks, and narrow the prompt before retrying after a safety stop.

    Tip: For long-running agent tasks, clearly define report ownership and handoff steps.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • RetiredSelf-hosted runner

    Failed runner debug logs are no longer uploaded automatically

    The self-hosted runner no longer uploads the debug log at the end of a session; logs from failed sessions remain on the runner host and are removed only with --remove-session-state.

    Usage details

    Who can use it: Self-hosted runner administrators.

    How to use it: Collect local runner logs; if removal is needed, explicitly provide the --remove-session-state option.

    Tip: Update your access and retention policy for local debug logs.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • AvailabilityClaude in Chrome / Remote Control / Claude iOS

    State synchronization improved in Chrome, Remote Control, and iOS clients

    Claude in Chrome waits six seconds for its browser to recover from brief disconnections and keeps javascript_tool output of up to 60.000 characters in the conversation. In Remote Control, the limit removed on an external device is refreshed; Claude iOS preserves the file card in single-file responses.

    Usage details

    Who can use it: Claude in Chrome, Remote Control, and Claude iOS users.

    How to use it: Update the clients and retry brief browser disconnections, limit removal, and single-file response flows.

    Tip: Continue to limit long JavaScript output with targeted queries.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • SettingClaude Code cloud

    Cloud permission sources and worker transitions are explained

    Cloud /permissions now says where rules are stored and how to change them. When a session switches workers, allow rules that cannot be preserved alongside deny/ask rules are not silently carried over and are requested again if needed; hidden slash commands also no longer appear in the mobile/web menu.

    Usage details

    Who can use it: Claude Code cloud sessions and web/mobile interfaces.

    How to use it: Run /permissions; review prompts that reappear after a worker transition and approve only what is needed.

    Tip: Define critical rules in organization-managed settings rather than at the session level.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • AvailabilityClaude Code cloud

    Cloud loop and task lifecycle are clearer and more robust

    Cloud /loop confirmation explains that the timer stops with the container and that a persistent option is available. Running tasks remain listed after /clear, Send now does not stop background agents, the memory connector is protected against name conflicts, and hot-reloading the mode no longer shuts it down during consecutive messages.

    Usage details

    Who can use it: Claude Code cloud sessions, scheduled tasks, and background agents.

    How to use it: In Cloud, use /loop, /clear, Send now, or the scheduled task flow; check the confirmation message, /tasks list, and connector name.

    Tip: For long-running work independent of the container, choose the persistent loop option.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • SettingCode Review / Claude Tag

    Code Review and Claude Tag admin conflicts are safer

    In Code Review, unlinking a GitHub organization also disables All repositories, which stays disabled until the organization is reconnected. Claude Tag setup can now start even if the channel list loads incompletely; an auto mode allow rule is rejected with a conflict message instead of silently overwriting another admin’s change.

    Usage details

    Who can use it: Code Review and Claude Enterprise Claude Tag administrators.

    How to use it: Retry the connection or setup; if there is a conflict, refresh the settings and save with the latest value.

    Tip: After reconnecting, review repository scope and concurrent admin changes.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • CapabilityClaude Code

    Config and file-saving workflows are more resilient

    Recent settings edits are no longer missed; if an old named config file exists, the active file is reported in stderr. The risk of files being left empty or truncated is reduced when the disk is full, files are read-only or bind-mounted, or another program has a config file open on Windows.

    Usage details

    Who can use it: All Claude Code installations, especially Windows, read-only, and bind-mounted workspaces.

    How to use it: Change the setting and verify it with /config; check the stderr warning and fix disk or write permissions.

    Tip: Back up your config and transcript before upgrading.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • Plugin / IntegrationClaude Desktop

    Desktop connector permissions and Bedrock tool list are fixed

    Rules written with a claude.ai connector terminal name, such as deny/ask/hook/disallowedTools, are no longer ignored in Claude Desktop. Gateway setups whose upstreams are all Amazon Bedrock no longer offer a web search tool that always fails.

    Usage details

    Who can use it: Organizations using connectors in Claude Desktop or only Bedrock upstreams.

    How to use it: Update Desktop and reload connector permission rules and the tool list.

    Tip: Copy connector names exactly into the rules and match tools to upstream capabilities.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • CapabilityClaude Code

    Diff and context display are more efficient in network projects

    /diff no longer breaks on files/folders named HEAD or after more than 64 edit turns. /context no longer resends the full visual grid on subsequent turns; terminal lists are reduced to the top 10 largest lines, and uncalled MCP tools to one line per server. Projects on Linux network file systems are also no longer scanned unnecessarily at startup.

    Usage details

    Who can use it: Claude Code terminal, claude.ai/mobile sessions, and Linux projects using NFS/SMB/Lustre.

    How to use it: Update to 2.1.319 and use /diff and /context as usual.

    Tip: In large sessions, check context and diffs regularly instead of relying on a single huge output.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • CapabilityClaude Code

    File editing, notebooks, and rewind are more reliable

    The Write tool no longer presents read errors as refusals; output from fast Bash jobs and the latest diff file are preserved; and editing the wrong cell in a notebook that changes during a permission prompt is prevented. /rewind can restore checkpoints from before a branch, and cannot restore or delete files outside the workspace through a symbolic link.

    Usage details

    Who can use it: Sessions using Claude Code Write, Bash, NotebookEdit, and /rewind.

    How to use it: Update and use the relevant tool as usual; verify write errors, the checkpoint, and the symbolic link target.

    Tip: Before critical rewinds, check your git commit and workspace boundaries.

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

    Gateway spend-limit message provides the reset date

    The spend limit error now shows the renewal date in the format spend limit reached · resets Dec 1 and includes the admin.blocked_message text, if configured.

    Usage details

    Who can use it: Developers using Claude apps gateway.

    How to use it: Read the message when you hit the limit; configure admin.blocked_message if needed.

    Tip: Use the reset date shown in this message in support responses.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • SettingClaude apps gateway

    Gateway reports policy and upstream issues at startup

    Gateway now warns at startup about conflicting custom model IDs, legacy Desktop settings used without desktopHome, mixing AWS GovCloud with other upstreams, and base URL variables in the environment. The Bedrock guardrail header not forwarded in managed settings and the upstream name for failed calls are also shown more clearly in logs.

    Usage details

    Who can use it: Claude apps gateway administrators; organizations using multiple upstreams, Bedrock, and Claude Desktop.

    How to use it: Restart Gateway and check the startup logs; adjust policies, upstream separation, base URLs, and the guardrail to comply.

    Tip: Send production gateway logs to a centralized collection system.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • AvailabilityClaude Code

    Sign-in source, billing organization, and expired sessions are more visible

    If a claude.ai session ends and you continue with an ant auth login identity, Claude Code now warns you and names the Claude Console organization billed at startup. On expired gateway sign-in, rejected telemetry/managed-settings requests are no longer retried; a gateway exported from the shell is also preserved after a background session update.

    Usage details

    Who can use it: Claude Code sessions using a Claude Console organization or Claude apps gateway.

    How to use it: Start Claude Code and check the sign-in source and organization warning; if the token expires, sign in again and keep the gateway environment variables in the process.

    Tip: Distinguish subscription usage from Console billing in operational documentation.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • SettingClaude Code

    Permission rules are more consistent with timeout, which, and deny/ask

    Read-only commands wrapped with timeout no longer request unnecessary permissions; unusual which flags do request permission, and a broad deny rule is applied before a narrower ask rule.

    Usage details

    Who can use it: Sessions using Bash/PowerShell permission rules, accept-edits, auto mode, and Monitor.

    How to use it: Review /permissions rules and test with small commands containing timeout and which; check the order of deny and ask rules.

    Tip: Narrow the scope of broad allow/bypass rules before using them in production.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • APIClaude API / Managed Agents

    Managed Agents onboarding templates updated

    Performance tuner has been removed from the /claude-api managed-agents-onboard templates, two templates have been renamed, and web search/fetch have been disabled in five templates that do not use these tools. The /claude-api skill also accurately explains how Console web settings affect Managed Agents and provides identity/permission examples.

    Usage details

    Who can use it: Managed Agents onboarding and Claude API developers.

    How to use it: Reload the command/skill and compare the new templates with existing agent permissions.

    Tip: Review template differences before automatically overwriting existing agents.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • APIClaude Code / Agent SDK

    MCP resume, SDK errors, and hook flow are more deterministic

    With an unavailable MCP server, claude -p --resume now returns exit code 1 and preserves the pending call. SDK informational text is no longer truncated, and disk-full errors return the actual error; PostToolUse hooks run for tool calls completed after an interruption and wait for up to 10 seconds.

    Usage details

    Who can use it: Integrations using MCP, claude -p, Agent SDK, and PostToolUse hooks.

    How to use it: Update the SDK/CLI; handle exit code 1 in CI, keep the MCP server accessible, and ensure hooks finish within 10 seconds.

    Tip: Design hooks to be idempotent and log preserved tool calls.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • SettingMonitor / Claude Code

    Monitor and proxy setting boundaries are clearer

    Monitor follows deny rules for every Bash command; with an ask rule, it does not skip the permission prompt for Monitor/bypass/auto mode. CLAUDE_CODE_PROXY_RESOLVES_HOSTS cannot be changed from a settings file on a host-managed network, and Monitor reconnects when using a WebSocket proxy with DNS conditions.

    Usage details

    Who can use it: Claude Code installations using Monitor and enterprise proxies.

    How to use it: Define rules and variables in the process environment, then test deny/ask and WebSocket flows.

    Tip: Document responsibility for network ownership and application permissions in a single configuration.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • APIClaude Code

    OpenTelemetry and proxy/TLS behavior are adapted for enterprise environments

    CLAUDE_CODE_CLIENT_CERT is now presented to certificate-requiring collectors with detailed tracing, and Agent SDK prewarm processes are no longer incorrectly counted in the session.count metric. The proxy DNS resolution setting cannot be changed from a managed file; TLS recovery no longer works with ALL_PROXY alone, requires HTTPS_PROXY, and can perform a secure TLS handshake when nonessential traffic is disabled.

    Usage details

    Who can use it: Organizations using OpenTelemetry, Agent SDK, and TLS/DNS proxies.

    How to use it: Set the required proxy variables and client certificate in the process environment; restart tracing and update the metric baseline.

    Tip: Centrally document the roles of ALL_PROXY/HTTPS_PROXY, the certificate chain, and network ownership.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • Plugin / IntegrationClaude Code

    Plugin, skill, and marketplace flows are safer and more consistent

    Plugin eval no longer suggests an invalid max: 0; automatic and manual plugin updates install new dependencies. The plugin subagent status line can use CLAUDE_PLUGIN_ROOT, skill hot-reload tracks deleted files, --add-dir registers the marketplace after trust approval, and under the sandbox, subfolder skills load only when explicitly entered.

    Usage details

    Who can use it: Plugin/skill developers and Claude Code projects using marketplaces.

    How to use it: Update the plugin and install its dependencies; accept the --add-dir trust dialog and, when needed, enter the skill directory with /cd.

    Tip: Use CLAUDE_PLUGIN_ROOT instead of a fixed absolute path for plugin script paths.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • AvailabilitySelf-hosted runner

    Self-hosted runner startup errors are reported earlier

    If the startup check for --use-anthropic-git-proxy or --configure-git fails, the runner now shuts down before accepting work. If model access cannot be refreshed, it retries earlier and writes a warning to the session; a stalled setup connection is also shown with the live/default environment name.

    Usage details

    Who can use it: Organizations using self-hosted runners.

    How to use it: Start the runner with the relevant options and check startup and model access logs.

    Tip: Verify runner startup checks and repo trust before checking the registration list.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • CapabilityClaude Code for VS Code

    VS Code chat panel is more resilient with long and accessible inputs

    Long slash-directory drag-and-drop items no longer lock up the panel; dictation works only in the active view and stops on the second press. Unclosed long math fences no longer hurt streaming performance; the diagnostics tool returns only problematic files in a large workspace, using compact JSON and an optional severity threshold.

    Usage details

    Who can use it: Claude Code for VS Code users.

    How to use it: Update the extension and try drag-and-drop, claude-vscode.toggleDictation, and diagnostics workflows.

    Tip: Use a minimum severity level for large diagnostic outputs.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • CapabilityClaude Code cloud / VS Code

    VS Code Web sessions no longer restore deleted turns or run unapproved plans

    The Claude Code for VS Code Web tab no longer restores turns removed with /clear, Retry, or Edit. A deleted session file is not restored by renaming it, and plan mode does not run without approval after a window refresh.

    Usage details

    Who can use it: Claude Code cloud web/mobile and Claude Code for VS Code Web sessions.

    How to use it: Update and test the /clear, Retry/Edit, session deletion, and plan refresh flows; read the warning when restoring.

    Tip: Save prompts you want to keep separately, and do not treat a refresh as plan approval.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • Plugin / IntegrationClaude Code

    Artifact database calls handle shared files more safely

    Calls to artifacts without an artifact database are no longer repeated, reads from out_dir are no longer rejected, and errors for missing documents or failed writes are explained. Shared writes use a hidden per-file marker; older or unmarked files are preserved and reported.

    Usage details

    Who can use it: Claude Code sessions using the Artifact database and shared files.

    How to use it: Retry the artifact call or out_dir read; inspect any existing unmarked file and the guidance in the error message.

    Tip: Deliberately merge older exports instead of automatically overwriting them.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • FeatureClaude Code

    CLI commands, login, and feedback flow behave more reliably

    Slash commands preceded by spaces, tabs, or line breaks now work; repeated 401 requests no longer refresh the login for other sessions, and the /login success screen closes without waiting. Surveys and /feedback drafts that cannot be submitted are no longer shown when an organization rejects feedback or a third-party gateway is used.

    Usage details

    Who can use it: Claude Code terminal, concurrent sessions, and enterprise/third-party gateways.

    How to use it: Update to 2.1.319 and try a slash command, /login, and, if needed, the /feedback flow; handle 401 exits separately in automation.

    Tip: Monitor the ready signal instead of using a fixed wait after login.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • Plugin / IntegrationClaude Code for VS Code

    Marketplace removal confirmation shows installed plugins

    Before removing a plugin marketplace, VS Code confirmation now checks and lists the plugins currently installed.

    Usage details

    Who can use it: Claude Code for VS Code marketplace users.

    How to use it: Start removing the marketplace and review the plugin list on the confirmation screen.

    Tip: Don’t remove the marketplace before moving any plugins you need.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları
  • FeatureClaude Code

    Model selection and file path error messages are more helpful

    /model groups older models under More models; for a misspelled model name, it suggests a likely correction instead of an update. On macOS/Windows, an external path that differs only in letter case is explicitly named in the refusal message, and the contrast of secondary terminal text is increased.

    Usage details

    Who can use it: Claude Code terminal users.

    How to use it: Open /model and use More models; verify the suggested model name and correct the spelling in the path refusal message.

    Tip: In automations, select model IDs from a fixed list.

    ✓ Verified by official sourceSource: Claude Code v2.1.319 sürüm notları

OpenAI · ChatGPT

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

No notable change since the last edition.

Meta · Meta AI

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

New in this edition · 1

  • OtherMeta Commercial Terms ve ilgili reklam şartları

    Meta pauses planned commercial terms updates

    Meta has paused planned updates to Meta Commercial Terms, Product Catalog Terms, Self-Serve Ad Terms, and Ad Creative Generative AI Terms. These changes will not take effect on November 20, 2026; the current terms remain in effect until a new announcement.

    Usage details

    Who can use it: Businesses using Meta’s advertising, catalog, and related commercial products; the current terms remain in effect, and the planned November 20 changes are not being implemented.

    How to use it: No feature is enabled and no API changes. Do not revise your compliance documents based on the planned new text; check the current terms in the Meta for Business Help Center.

    Tip: Keep compliance and contract reviews aligned with the current terms, and watch for Meta’s next announcement.

Moonshot AI · Kimi

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

New in this edition · 7

  • RetiredKimi Code server API

    Kimi Code removed legacy REST transcript endpoints

    The `GET /sessions/{id}/transcript`, `/transcript/ops`, `/transcript/plan`, and `/transcript/user-messages` endpoints are being removed in the new conversation protocol. They are replaced for external clients by an agent-based turns retrieval endpoint and new WebSocket events.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; external clients that directly use REST transcripts are affected.

    How to use it: Replace client code that calls the legacy transcript endpoints with `GET /sessions/{id}/agents/{agent_id}/turns`; implement `subscribe_v2` and `turn.*`/`item.*` events for live updates.

    Tip: If you need to support both old and new server versions during migration, first check the server’s live OpenAPI/AsyncAPI schema and detect the protocol version.

  • APIKimi Code server API

    Kimi Code serves the conversation stream as turns and items instead of a transcript

    The Kimi Code server now produces and streams conversations as turn and item objects instead of a single transcript projection. The WebSocket side uses `turns`, `turn.*`, `item.*`, and `notice` frames, while conversation history is served through an agent-based turns endpoint.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; not included in the stable 2.5.2 release notes.

    How to use it: Use the new `subscribe_v2` payload (`session_id`, optional `agent_ids`) when subscribing to the WebSocket; process the `turns` page and `turn.*`/`item.*` events. Read history through `GET /sessions/{id}/agents/{agent_id}/turns`.

    Tip: Do not treat subscription acknowledgment as meaning that all data has arrived; handle the initial `turns` page and subsequent events separately, then merge them by turn ID.

  • Plugin / IntegrationKimi Code CLI

    Custom registry can now import Google GenAI providers

    Entries with `type: google-genai` in `api.json` can now be imported using `/provider` or `kimi provider`; registries containing only this type no longer fail either. The base URL is normalized to prevent duplication of the `/v1beta` path added by the Google GenAI SDK or the `/v1beta1` path for Vertex.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; Kimi Code CLI users with a Google GenAI-compatible custom registry.

    How to use it: Open `/provider` in the TUI or `kimi provider` in the shell and import the custom registry `api.json` URL; leave the provider type as `google-genai`. Prefer specifying only the gateway root for `base_url`; entries that include a version path are no longer duplicated either.

    Tip: Explicitly define the credential for your custom Gemini gateway using `api_key` or `api_key_env`, then verify the catalog after importing with `/provider refresh <providerId>`.

  • CapabilityKimi Code CLI

    Updated tool schemas are used even when dynamic loading is disabled

    Kimi Code no longer uses the stale startup schema in requests when a tool or MCP definition with the same name changes. The fix selects the current definition both in sessions that use dynamic tool loading and for models or requests that do not support dynamic loading; new runtime tools continue to follow the existing loading rules.

    Usage details

    Who can use it: 2.6.0-canary.20261106040415-d37c9d2 and later canary builds; CLI, SDK, Desktop, Web, and ACP sessions using updated MCP/inline tools.

    How to use it: Update Kimi Code to this canary version; run it in the next turn after changing a tool or MCP schema. For a custom registry catalog, use `/provider refresh`; for MCP changes, use the existing MCP refresh/authentication flow.

    Tip: When tool parameters change, verify with a test call that the model sees the current schema before reusing an old session.

  • APIKimi Code server API

    Kimi Code removed the source field from search responses

    Session search responses no longer include the `source` field indicating whether a result comes from `live` or `index`. Session-scoped search also runs against the persistent index; text that has not yet been indexed may not be searchable.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; clients using `POST /search`.

    How to use it: Remove branches that depend on the `source` field from your API client and process only the returned search results. To have past sessions indexed, use the `[database] global_search = true` or `KIMI_CODE_GLOBAL_SEARCH=1` setting.

    Tip: If you need to show a source label on the client side, derive it from your own search mode rather than relying on a server field, and do not treat its absence as an error.

  • APIKimi Code server API

    Skill activation now accepts an optional prompt_id

    Kimi Code’s skill activation API now accepts the `prompt_id` field. This ID participates in client-side matching and duplicate request prevention in the same way as regular prompt submissions.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; clients that activate skills using the API or SDK.

    How to use it: Add a client-generated unique `prompt_id` to the body of the `POST /sessions/{id}/skills/{name}:activate` request and use the same ID for UI bubble or retry matching.

    Tip: For network retries, keep the same `prompt_id` value instead of generating a new ID for each activation to reduce the risk of duplicate activations.

    ✓ Verified by official sourceSource: Kimi Code PR #4663 API protokolü
  • CapabilityKimi Code Web

    Streamed thinking and response items now end only once

    In the Kimi Code Web stream, a thinking or response item ends with just one `item.ended` event once its content is committed to the persistent record. Duration now comes from the measurement recorded by the engine rather than the client stream clock, so the thinking duration no longer jumps slightly afterward.

    Usage details

    Who can use it: 2.6.0-canary.20261106115820-bc6fa89 and later canary builds; users of the Kimi Code Web/Desktop stream.

    How to use it: On the client side, handle an item’s end based on the single closing event from the persistent record, not the first early `item.ended` event; no separate setting is required.

    Tip: Treat `duration_ms` as the source of truth instead of recalculating the thinking duration in the stream interface.

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 1

  • AvailabilityGrok API

    Output Watermarking is enabled by default on supported models

    Starting November 7, 2026, text outputs from supported models include Output Watermarking by default, with no request changes required. This activation covers regular text, structured outputs, and tool-call arguments in the Responses API and Chat Completions API.

    Usage details

    Who can use it: Applies to xAI API teams on global and regional endpoints; covers grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the three grok-4.20 models. Team administrators can disable it per model; some models may be locked for the team.

    How to use it: Read the x-watermark header in the API response without changing your code; a value of true indicates that the output is watermarked. As an administrator, open Console → Team settings → Model output watermarking, clear the model checkbox using the settings icon, then select Save → Opt Out.

    Tip: In production integrations, log the x-watermark header and report watermarked and unwatermarked outputs separately; changes apply to new requests within about 5 minutes.

    ✓ Verified by official sourceModels: grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning, grok-4.20-multi-agentSource: Output WatermarkingSource: TypeScript SDK değişiklik günlüğü

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 6

  • SettingLoopForge

    LoopForge added fixed network environment selection and key validation for CodeBuddy

    In Personal settings and Quick Access, the CodeBuddy network environment is no longer free text; International, China, and iOA are selected from a list. The API rejects undefined values, opens the key page based on the selection, and shows a non-blocking warning for incorrect keys such as DeepSeek.

    Usage details

    Who can use it: Merged into LoopForge’s main branch on November 7; applies to users configuring CodeBuddy credentials. No tagged release or deployment date is specified in the commit.

    How to use it: Open Personal settings or Quick Access → CodeBuddy credentials, select the environment where the key was created from the Network environment list, paste in the API Key, and save. Then use the key page that opens based on your selection.

    Tip: Select the same environment as the CodeBuddy site where you created the key; don’t paste third-party keys such as `sk-` into the CodeBuddy field.

  • CapabilityHyInv

    HyInv added support for PDB insertion codes with the --fixed option

    HyInv’s fixed-position PDB parser can now read and select insertion codes. `A12B` selects only residue 12B in chain A, while `A12.` selects residue 12 without an insertion code; range endpoints without a specified insertion code include the relevant insertions.

    Usage details

    Who can use it: HyInv source code in Tencent-Hunyuan/Hyra-results; for CLI and Python API users. The change was published in the November 6 commit; no separate package release is specified.

    How to use it: Install HyInv and run `python -m hyinv design --pdb structure.pdb --fixed A12A-12C --n 16`; use the chain prefix for multi-chain PDBs. In Python, you can pass `design(..., fixed="A12A-12C")`.

    Tip: The old `A12` syntax preserves all insertions at residue number 12; use `A12.` to preserve only the main residue without an insertion code.

  • Capabilityteamai-cli

    teamai-cli enabled judging workflows on every driver; the default is local-only

    Every driver can now run judging workflows; the default run scope is set to local-only. In addition to the known scorer-layer change, this opens judging workflows to all drivers and narrows the default scope.

    Usage details

    Who can use it: All drivers in the teamai-cli v0.30.0-beta.0 beta; team settings are required to switch to a shared run scope, but the exact setting name isn’t provided in the release notes.

    How to use it: Upgrade to the beta, run a judging workflow from your driver, and keep the scope at the local-only default or expand it according to your team policy. The release notes don’t specify the exact flag for changing the scope.

    Tip: Keep the local-only default during initial trials; enable shared judging runs only after reviewing their cost and data-access implications.

  • Plugin / Integrationteamai-cli

    teamai-cli sessions can now read Hermes coding transcripts

    The Sessions feature has been expanded to support reading Hermes coding transcripts. Hermes session records can now be used as a source in teamai-cli’s session-reading workflow.

    Usage details

    Who can use it: teamai-cli v0.30.0-beta.0; for teams using Hermes coding transcripts.

    How to use it: Upgrade to the beta and run teamai-cli’s session-reading or import workflow in the workspace containing the Hermes coding transcript. The release notes don’t specify a separate command or file path for this workflow.

    Tip: When moving transcript files, keep the directory structure and format generated by Hermes unchanged; otherwise, automatic discovery may not work.

  • Featureteamai-cli

    teamai-cli added copy ownership and run handoff flows to workflows

    A workflow can now take ownership of its own copy, assign a run to a person, clean up run branches, and read a step’s PR. This is a run lifecycle feature separate from the existing workflow/scorer architecture record in v0.30.0-beta.0.

    Usage details

    Who can use it: Teams using the teamai-cli `v0.30.0-beta.0` beta; the release notes don’t specify a plan, operating system, or general access limit.

    How to use it: Upgrade teamai-cli to `v0.30.0-beta.0` and run your workflow; use the ownership, assignment, run-branch cleanup, and step PR reading flows within the workflow run. The release notes don’t specify separate command names for these operations.

    Tip: For long-running jobs, define the handoff point and branch cleanup policy up front; use PR reading before verifying the handoff.

  • Settingteamai-cli

    teamai-cli now accepts the `agentsmd/` directory for shared instructions

    The `agentsmd/` directory has been added as a source of shared instructions. Teams can now keep shared instruction files in this directory and include them in teamai-cli’s shared-instructions workflow.

    Usage details

    Who can use it: teamai-cli v0.30.0-beta.0; for team repositories and workspaces that use shared instructions.

    How to use it: Upgrade to the beta, create an `agentsmd/` directory in the shared instruction source, add team instructions there, and rerun the workflow. The release notes don’t specify a filename or precedence rule.

    Tip: Keep general team rules under `agentsmd/` and project-specific exceptions in narrower-scope instruction files to reduce conflicts.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 35

  • APIQwen Code CLI

    Interactive CLI reports terminal status with OSC 7501

    The interactive Qwen Code CLI writes OSC 7501 status records to the TTY, reporting working, blocked, done, error, and idle states, along with permission, question, or authentication types. `-p`, JSON, stream-JSON, and ACP flows do not emit these records.

    Usage details

    Who can use it: Qwen Code Desktop v0.27.0 interactive CLI; terminals or agent inbox integrations that support OSC 7501.

    How to use it: Start interactive `qwen` in a TTY that supports OSC 7501, then run a tool that prompts for permission or a question.

    Tip: Visualize the status record in a terminal tab or agent inbox.

  • Plugin / IntegrationQwen Code

    OCI Generative AI added to OpenAI-compatible provider documentation

    Qwen Code documentation now lists Oracle Cloud Infrastructure Generative AI among OpenAI-compatible providers.

    Usage details

    Who can use it: Qwen Code nightly 2026-11-07; users with OCI Generative AI access.

    How to use it: Select the OpenAI-compatible path in the provider configuration and enter the official endpoint, model, and credentials for OCI.

    Tip: Match the OCI model name and base URL with the current values in the provider documentation.

  • APIQwen Code TypeScript SDK

    TypeScript SDK v0.1.30 released, bundling CLI 0.27.0

    TypeScript SDK v0.1.30 has been released, and the official listing says this version bundles CLI source version 0.27.0.

    Usage details

    Who can use it: TypeScript SDK users.

    How to use it: Upgrade the SDK dependency to v0.1.30 and verify that the bundled CLI version is 0.27.0.

    Tip: Use the SDK and CLI together as the 0.1.30/0.27.0 pair in this release.

  • CapabilityQwen Code CLI

    Startup context now includes the active shell name

    The active shell name is now reported in the Qwen Code CLI startup context, helping distinguish between different shells or host environments.

    Usage details

    Who can use it: Qwen Code nightly 2026-11-07.

    How to use it: Start the nightly CLI as usual; the active shell name is added automatically.

    Tip: Check the startup context when testing the same project in different shells.

  • Plugin / IntegrationQwen Code

    Wally documented in OpenAI-compatible provider documentation

    Wally has been added to the OpenAI-compatible provider documentation in the Qwen Code nightly build. No separate model or pricing announcement was made.

    Usage details

    Who can use it: Qwen Code nightly 2026-11-07; developers using the Wally endpoint.

    How to use it: Open the OpenAI-compatible provider section and enter the official endpoint, model, and key values for Wally.

    Tip: Match the Wally endpoint and model names with the official documentation.

  • Plugin / IntegrationQwen Code Web Shell

    Repository allowlist and blocklist filters added to GitHub Channel

    GitHub Channel notifications can now be filtered by repository using an allowlist or blocklist. Filtered notifications are recorded in the timeline as `repo_filter`.

    Usage details

    Who can use it: Web Shell → GitHub Channel; also applies to the daemon API, default Off.

    How to use it: Go to Settings → Channels → GitHub, set Repository Filter to Allowlist or Blocklist, and add `owner/repo` or GitHub URLs to Filtered Repositories.

    Tip: Add noisy repositories to the blocklist; note that when used with Group Policy, the effective set is the intersection.

  • AgentQwen Code

    Multiple child questions can now be pending simultaneously in Managed Agent sessions

    Managed Agent sessions can now hold multiple pending child user questions at the same time. Questions can be answered in any order; drafts and errors are kept separately for each child question.

    Usage details

    Who can use it: Qwen Code v0.27.0 Managed Agent sessions; plan and region not specified.

    How to use it: When multiple Ask User Question outputs appear, select and answer any of the pending cards.

    Tip: If you’re not ready to answer one question, answer the other child questions to keep the session moving.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • FeatureQwen Code Web Shell

    Web Shell adds a Q&A activity heatmap and usage view for long turns

    During a long-running turn, the waiting card shows 44 weeks of daily Q&A activity. The Activity/Usage toggle displays token details; clicking a day takes the transcript to that day’s first Q&A.

    Usage details

    Who can use it: Main Web Shell composer; hidden on coarse-pointer devices.

    How to use it: Send a prompt; while the response is streaming, open the heatmap button at the end of the composer status line and use the Activity/Usage views.

    Tip: Click the day with increased Q&A activity to return to the start of the decision in the transcript.

  • FeatureQwen Code Web Shell

    Ask for details side task added to pending user questions

    Pending Ask User Question panels can now open an Ask for details side task that explains the options. The side task does not answer the original question on the user’s behalf.

    Usage details

    Who can use it: Pending Ask User Question panels in Web Shell.

    How to use it: Select Ask for details on the pending card and read the explanation; then answer the original card as usual.

    Tip: Open the details side task first if the effects of the options are unclear.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • APIQwen Code

    Core API added to read a single PR’s status by URL for the daemon

    `fetchGitHubPullRequestView(url, cwd, env?)` was added to core. The API returns the PR title, head SHA, status, mergeability, CI summary, and failed checks; the refresh call will be wired up in a subsequent step.

    Usage details

    Who can use it: Qwen Code core/daemon integration developers; the API is available, but the PR refresh consumer is not yet connected.

    How to use it: Call `fetchGitHubPullRequestView` with a validated `https://<owner>/<repo>/pull/<n>` URL and the current working directory.

    Tip: Restrict external input to allowed GitHub hosts before making the call.

  • CapabilityQwen Code

    Interactive TUI surfaces approval requests from background workflow agents

    Interactive TUI shows when background workflow agents are waiting for tool approval and opens task details. Work awaiting approval no longer gets stuck silently.

    Usage details

    Who can use it: Qwen Code v0.27.0 interactive TUI and workflow agent runs.

    How to use it: Open the approval notification and approve the tool from the background-task details.

    Tip: Regularly check background approval notifications during long workflows.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    Cancelled shell commands are reported as interrupted and retain output

    Shell commands cancelled with Esc or Ctrl+C are now shown as interrupted instead of completed. Any partial output produced up to the moment of cancellation is preserved.

    Usage details

    Who can use it: Qwen Code v0.27.0 interactive shell runs.

    How to use it: Press Esc or Ctrl+C while a shell command is running; check the Interrupted status and partial output.

    Tip: Save partial output as debugging evidence.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    serve preserves capacity by reclaiming idle channel sessions

    serve reclaims idle sessions for cached channel routes; unused routes do not consume the maxSessions budget.

    Usage details

    Who can use it: Qwen Code v0.27.0 serve/channel runs.

    How to use it: Upgrade serve to v0.27.0; idle sessions are reclaimed automatically.

    Tip: Plan the maxSessions value based on active concurrent usage.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    Shell edits handle CRLF and Unicode safely in sed

    The Shell tool’s sed emulation correctly handles CRLF files and Unicode characters in substitution patterns. The risk of corruption with Windows line endings is reduced.

    Usage details

    Who can use it: Qwen Code v0.27.0 Shell tool.

    How to use it: Request a standard sed-based edit; no extra flags are needed.

    Tip: Edit Windows sources without converting line endings, and check the diff.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code Managed Agent

    Upgrade rejections and Advisor errors conceal embedded credentials

    Upgrade rejections display passwords in masked form; Advisor startup errors do not expose embedded credentials. Empty Harness user credentials are rejected during the start and verify stages.

    Usage details

    Who can use it: Qwen Code v0.27.0 Managed Agent upgrade, Advisor, and Harness start/verify flows.

    How to use it: Run upgrades or verification as usual; if an error occurs, use the sanitized message and fill in the Harness user credentials.

    Tip: Still review the sanitized message before sharing error logs.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • FeatureQwen Code Web Shell

    Web Shell shows tool arguments and results in expanded cards

    When Web Shell tool cards are expanded, you can see the arguments used alongside the tool result.

    Usage details

    Who can use it: Qwen Code v0.27.0 Web Shell tool cards.

    How to use it: Expand a tool call’s card and open the arguments and results sections.

    Tip: If a result is unexpected, check the arguments first.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    Anthropic-compatible provider refusals now show more detail

    Qwen Code now surfaces refusal details returned for Anthropic-compatible requests, making it possible to inspect the reason for an upstream refusal instead of seeing a generic failure.

    Usage details

    Who can use it: Qwen Code v0.27.0 Anthropic-compatible provider flows.

    How to use it: Run a request with an Anthropic-compatible provider; if it is refused, inspect the refusal details.

    Tip: Don’t assume a prompt or tool error before addressing the reason for the refusal.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • APIQwen Code serve

    Invalid start responses are now reported as definitive rejections

    The serve API classifies a malformed start response as a definitive rejection rather than a transient condition.

    Usage details

    Who can use it: Qwen Code v0.27.0 serve API.

    How to use it: When starting a session, send a response body that matches the expected start payload; if rejected, correct the payload.

    Tip: Treat a payload-invalid response as a permanent error rather than automatically retrying.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code Managed Agent

    Managed Agent session renaming errors fixed

    Renames made before an in-progress title delivery is complete no longer result in a 503 error.

    Usage details

    Who can use it: Qwen Code v0.27.0 Managed Agent sessions.

    How to use it: Rename a Managed Agent session as usual; while the operation is in progress, the system waits for title delivery.

    Tip: Don’t repeat the rename request in quick succession.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    Non-interactive and ACP runs warn about excessive Skill stacking

    Non-interactive, ACP, and channel runs issue a warning when more than five Skills are added to a prompt. The warning does not interrupt the model’s response.

    Usage details

    Who can use it: Qwen Code v0.27.0 non-interactive, ACP, and channel runs.

    How to use it: Check the warning for a prompt containing six or more Skills, then remove any unnecessary Skills.

    Tip: Stack only the Skills needed for each task.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • FeatureQwen Code

    Text-editing controls restored in OpenTUI dialogs

    OpenTUI dialogs have regained the text-editing controls that had disappeared.

    Usage details

    Who can use it: Qwen Code v0.27.0 OpenTUI interactive interface.

    How to use it: Open a dialog in the OpenTUI interactive flow and use the editing controls in the text field.

    Tip: For long confirmation text, check cursor behavior before submitting.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • AvailabilityQwen Code Web Shell

    Interfaces without a relay dictionary fall back to English text

    If no relay dictionary is available for the selected language, Web Shell uses English fallback text instead of blank or broken text.

    Usage details

    Who can use it: Web Shell interface languages with an incomplete relay dictionary.

    How to use it: Set the Web Shell language to one with an incomplete relay dictionary; the fallback text is used automatically.

    Tip: Refer to the English fallback text when a translation is missing.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • AvailabilityQwen Code

    Follow-up suggestions retain Thai, Lao, Khmer, and Myanmar

    The language of follow-up suggestions no longer unintentionally switches to English in Thai, Lao, Khmer, and Myanmar interfaces.

    Usage details

    Who can use it: Qwen Code v0.27.0 Thai, Lao, Khmer, and Myanmar language flows.

    How to use it: Set the interface to one of these languages and open the follow-up suggestions after a response.

    Tip: Check the terminology before using suggestions as your next prompt.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code

    In fresh-turn requests, the user’s instruction remains the final content

    In new-turn model requests, injected context is placed before the user’s instruction, preserving the final instruction as the last content.

    Usage details

    Who can use it: Qwen Code v0.27.0 model requests.

    How to use it: Send your instruction as usual in a new turn; additional context is automatically placed before it.

    Tip: State exceptions that should apply after the context clearly in the final instruction.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • SettingQwen Code VS Code companion

    VS Code provider settings preserve other settings.model fields

    When configuring a provider, the VS Code companion writes only the selected model name and baseUrl; sibling settings in settings.model are preserved.

    Usage details

    Who can use it: Qwen Code v0.27.0 VS Code companion.

    How to use it: Update the provider model or baseUrl in the VS Code companion.

    Tip: If you use multiple provider profiles, don’t delete custom model fields separately.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • CapabilityQwen Code Web Shell

    Web Shell tool cards are labeled with lifecycle toolName

    When a regular name is unavailable, Web Shell tool cards use the lifecycle toolName, reducing the number of Unknown Tool displays.

    Usage details

    Who can use it: Qwen Code v0.27.0 Web Shell tool cards.

    How to use it: Run a tool that carries lifecycle metadata and check that the result card is labeled with the actual toolName.

    Tip: Record the lifecycle name shown during diagnosis alongside the tool call.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • SettingQwen Code Web Shell

    Optional Tetris and Snake games can be enabled on the Web Shell waiting card

    The waiting card now has an option for Tetris or Snake, both off by default. The games appear only on hosts where they are explicitly configured.

    Usage details

    Who can use it: Web Shell hosts; opt in via URL, localStorage, or host injection.

    How to use it: Add `?waitingGames=1` to the URL; alternatively, use `localStorage['qwen-code-web-shell-waiting-games']='1'` or `window.__QWEN_CODE_WAITING_GAMES__ = true`.

    Tip: Keep the default-off behavior in production and enable it only in test environments.

  • FeatureQwen Code Web Shell

    The Web Shell release dialog also lists channel sessions

    The Web Shell release dialog now includes channel sessions, so sessions originating from channels no longer remain hidden during the update flow.

    Usage details

    Who can use it: Qwen Code v0.27.0 Web Shell release dialog.

    How to use it: Open the release or update dialog in Web Shell and check the list for channel sessions.

    Tip: Verify active delivery flows in the list before updating.

    ✓ Verified by official sourceSource: Qwen Code v0.27.0 sürüm notları
  • AgentAlibaba Cloud SLS DataAgent

    SLS DataAgent turns natural-language questions into verifiable business outcomes

    Using real-time operational data and a Context Model, SLS DataAgent turns natural-language questions into query plans, analysis results, and verifiable business outcomes.

    Usage details

    Who can use it: Business users of SLS DataAgent; plan, country, and region not specified.

    How to use it: Open the DataAgent linked to the relevant Context Model, ask a business question in natural language, and review the query plan and evidence.

    Tip: Check the Context Model and evidence chain before taking critical actions.

  • CapabilityAlibaba Cloud Simple Log Service (SLS)

    SLS adds a semantic layer that brings business context to Agents

    The SLS semantic layer standardizes the meaning of fields, metrics, and terms by creating a Source Model and Context Model. DataAgent, enterprise Agent APIs, and copilot tools can all consume the same context.

    Usage details

    Who can use it: Enterprise teams using business data and Agent integrations on SLS.

    How to use it: Create a Source Model from SLS data assets, select a Context Model based on the business domain, and connect it to DataAgent or API calls.

    Tip: Use a shared Context Model instead of defining metrics separately for each Agent.

  • FeatureAlibaba Cloud Simple Log Service (SLS)

    SLS launches SessionStore for Agent workflows

    SLS SessionStore organizes and reconstructs Agent messages, model calls, and tool calls at the session level. It produces trajectory, session, and feature views through managed direct write or Ref association.

    Usage details

    Who can use it: Alibaba Cloud SLS users and enterprise teams collecting Agent operational data; plan or region not specified.

    How to use it: Send Agent data using managed direct write or connect a LogStore with Ref association; configure a mapping profile/schema profiler and query the views with SQL.

    Tip: Use a single mapping profile to produce troubleshooting, evaluation, and cost analysis views together.

  • Plugin / IntegrationAlibaba Cloud SLS MCP Server

    SLS releases an MCP Server with OAuth and RAM controls

    SLS MCP Server lets MCP-compatible Agents use resource discovery, log analysis, and alert diagnosis through a standard protocol. OAuth, a local proxy, and RAM permissions reduce the need to manage AccessKeys.

    Usage details

    Who can use it: MCP-compatible Agents and enterprise teams using SLS.

    How to use it: Add the SLS MCP Server connection to an MCP-compatible Agent, sign in with OAuth, and make calls within RAM permissions.

    Tip: Use OAuth and the narrowest RAM permission set instead of putting an AccessKey in the prompt.

  • Plugin / IntegrationAlibaba Cloud SLS Skills framework

    SLS releases an orchestration-ready Skills framework

    SLS released a framework that connects specialized Skills through the `alibabacloud-sls-agent-workflow` orchestration Skill. Users provide goals in natural language; subtask results are collected with status, evidence, and sources.

    Usage details

    Who can use it: Skill and Agent orchestrations running with SLS.

    How to use it: Add the `alibabacloud-sls-agent-workflow` Skill to the Agent environment, provide a goal in natural language, and review the results from the sub-Skills.

    Tip: Keep security gates and approval steps in place before production changes.

  • CapabilityQoder CLI

    Qoder CLI fixes the pending authentication status in `/mcp` listings

    Qoder CLI 1.1.84 correctly displays the status of MCP services awaiting authentication in the `/mcp` interface.

    Usage details

    Who can use it: Qoder CLI 1.1.84 and MCP services.

    How to use it: Upgrade to 1.1.84, run `/mcp`, and track the pending authentication status.

    Tip: If you have connection issues, check the `/mcp` status and pending authentication before reconfiguring.

    ✓ Verified by official sourceSource: Qoder CLI resmi changelog, 1.1.84
  • CapabilityQoder CLI

    Qoder CLI preserves recent responses after restart and resume

    Qoder CLI 1.1.84 improved compaction retention so recent responses and tool results are not lost when a session is restarted and resumed.

    Usage details

    Who can use it: Qoder CLI 1.1.84; plan, country, and operating system not specified.

    How to use it: Update to 1.1.84; after a long session, use `/resume` and check the response and tool records.

    Tip: Save critical checkpoints separately during long sessions.

    ✓ Verified by official sourceSource: Qoder CLI resmi changelog, 1.1.84

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 on October 7; a fast and inexpensive option at $0.10 / $0.50.

Grok 4.7

35
xAI · Grok

The default model in the Grok app. Since November 7, text outputs from supported models in the Grok API have been watermarked by default (`x-watermark` header).

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 has been increased tenfold (not officially confirmed).

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba’s flagship cloud model; Qwen 4 has not yet been released. Qwen Code v0.27.0 added concurrent sub-questions, a GitHub repository filter, and status reporting via OSC 7501.

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