YZ·RADARyz.mikail.net
Daily edition · 6 November 2026

GPT-6 Cyber launches, Claude gets Personal Finance, Gemini Omni 1.2 GA

No changes to the TOP 10 today either; the ranking and scores of all ten models are the same as yesterday. Gemini 4 Carbon leads with 60 points, followed by Claude Opus 5.5 (59) and GPT-6 Astra (57). Artificial Analysis lists the index version as v5.0.1. The watchlist is also unchanged; we’re only correcting yesterday’s note about Tencent Hy4: the model is now available via API through OpenRouter (`tencent/hy4`), discounted by 50% on Tencent Cloud TokenHub through December 31, while WorkBuddy and CodeBuddy are free through November 15. The LMArena page was again blocked by CAPTCHA; Gemini 4 Carbon also tops the accessible view dated October 27.

OpenAI announced GPT-6 Cyber for cybersecurity research. The model can find and verify vulnerabilities, and is designed to stay within the scope set by the researcher; OpenAI reports a 51% relative improvement over GPT-5.6 Cyber on ExploitGym. Access is currently limited to approved Daybreak Red customers via API and Codex. The legacy `gpt-5.5-cyber-preview` will be removed from the API on February 1, 2027. Python 3.46, Node 7.51, and Go 3.95 SDKs for developers added data residency options in Australia, India, and Japan.

Anthropic brought the Personal Finance beta to Claude in the US for Pro, Max, and Ultra plans. Balances for bank, card, and investment accounts connected through Plaid appear in one place; access is read-only. A security package arrived for enterprise customers: IP allowlist (beta, up to 1,000 entries per organization), using the Admin API to sign a member out of all sessions, Security Center to compare settings against recommendations, multi-organization management at admin.claude.com, Claude Desktop access controls, and an EDR sensor for the Cowork virtual machine. At Google, Gemini Omni 1.2 Flash became generally available on the API and Agent Platform. It generates videos up to 15 seconds long and accepts six reference images; rate limits doubled on the paid tier, and video output is 30% cheaper. The model also came to Vids and is rolling out to Ultra subscribers in the Gemini app. Requests for `gemini-omni-1.1-flash` will be routed to 1.2 on December 3. The Gemini API can now read PPT, ODP, ODS, and older DOC files, and Gemini-powered response analysis (Actionable Insights) came to Google Forms.

Meta added real-time voice calls, hundreds of voice options, and avatars that react to conversations to the Muse app in the US. WhatsApp Events Home brings event and RSVP management together in one place; invitations can also be sent via a link to people who don’t use WhatsApp. Graph API v21.0 will be shut down on January 21, 2027, and v22.0 on May 20, 2027. Kimi Work 3.2.25 added global search and “Stop and Archive”; 32 entries were added alongside Kimi Code 2.5 canary releases. With Qwen Code v0.26.8, users can sign in with a Qwen subscription in the CLI, VS Code, and Desktop; project hooks now require approval (42 entries for Qwen). Qwen is being tested within Apple Intelligence in the iOS 27.3 developer beta in China. xAI, DeepSeek, Z.ai, and Xiaomi had no new user-facing changes during this period. A total of 124 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

    Announced on October 28; tops the list 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. Personal Finance beta arrived in Claude in the US; Security Center and IP allowlist opened for Enterprise.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. GPT-6 Cyber from the same family is available to approved Daybreak Red customers via API and Codex.Released: 3 September 2026

  4. 04

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

    The default model for ChatGPT paid 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 the claude.ai free 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 API; focused on vision and computer use. Dots agents are rolling out to Enterprise, Edu, and Healthcare (admin preview through November 18).Released: 26 October 2026

  7. 07

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available on Google AI Pro/Ultra and paid API since October 13; $2 / $10 introductory pricing through November 30.Released: 30 September 2026

  8. 08

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta’s most powerful Muse model; no public API. Real-time voice calls and expressive avatars came to the Muse app in the US.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

    Released on October 23; API $2 / $10, open weights on Hugging Face (2.8T total / 104B active, 1M context). The only open-weight model in the TOP 10.Released: 23 October 2026

Key signals

  • No changes to the TOP 10 rankings or scores: Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57. The watchlist is also unchanged; Tencent Hy4 is now available via API on OpenRouter.
  • OpenAI: GPT-6 Cyber is available to approved Daybreak Red customers through API and Codex (51% improvement over GPT-5.6 Cyber on ExploitGym). gpt-5.5-cyber-preview will be shut down on February 1, 2027; SDKs added data residency in AU/IN/JP.
  • Anthropic: Personal Finance beta in Claude (US, Pro/Max/Ultra, read-only via Plaid). Enterprise gets IP allowlist, Sign out everywhere, Security Center, a centralized admin console, Desktop controls, and a Cowork EDR sensor.
  • Google: Gemini Omni 1.2 Flash GA (15-second video, 6 references, 30% cheaper video output); 3–15-second clips in Vids. Omni 1.1 routes to 1.2 on December 3. Gemini API reads PPT/ODP/ODS/DOC; Actionable Insights came to Forms.
  • Meta: real-time voice calls and expressive avatars in Muse (US). WhatsApp Events Home and invitations for guests without accounts. Graph API v21/v22 shutdown dates set for 2027.
  • Kimi Work adds global search and Stop and Archive; Qwen Code v0.26.8 adds subscription sign-in and project hook approval; Hy4 is 50% off on TokenHub. No new changes at xAI, DeepSeek, Z.ai, or Xiaomi.

Worth trying today

01

Have Gemini analyze your form responses

Open a form in Google Forms and go to the `Responses` tab; click `Generate insights` in the `Actionable Insights` section. Gemini runs Python on the responses to surface trends, relationships between questions, and suggested next steps; you can check the calculations in `Analysis steps`. It’s useful for a quick first look at education evaluation forms, but interpret suggestions cautiously when the sample is small. Rolling out November 2 on Rapid Release and November 19 on Scheduled Release.

Google · Gemini
02

Review your Claude organization with Security Center

In Claude Enterprise, go to Organization settings > Security posture and start with checks marked Failed and Not attested. Then, under Organization and access, enable IP allowlist and add your office, VPN, and CI egress addresses (changes can take up to 15 minutes to apply). If a device is lost, use “Sign out everywhere” on the Members page to end all of that person’s sessions in one step.

Anthropic · Claude
03

Start with low-resolution drafts in Gemini Omni 1.2

Select the `gemini-omni-1.2-flash` model in AI Studio. First generate short drafts at 360p or 720p, then upscale the best one to 1080p or 4K; this lowers the cost of re-generating. You can use up to six reference images for the same product or character. If your code still uses `gemini-omni-1.1-flash`, change the model ID instead of waiting for automatic routing on December 3.

Google · Gemini
04

Sign in to Qwen Code with your subscription

Update Qwen Code to v0.26.8. In the CLI, select “Qwen Account with Subscription” from the `/auth` menu; in VS Code and Desktop, select Qwen from the provider list and complete authorization in your browser. If you run out of quota, use the `/usage` and `/upgrade` commands. In the new version, project hooks require approval before running; reject the request in repositories you don’t recognize.

Alibaba · Qwen

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 · 11

  • Pricing / PlanGemini Enterprise

    Prepaid credits and auto-reload reach GA in Gemini Enterprise

    Self-serve online Cloud Billing accounts can now purchase prepaid credits for Gemini Enterprise Pay-as-you-go and over-quota usage. Credit balances can be used across Gemini Enterprise apps, Gemini Enterprise Agent Platform, and AI developer tools such as Antigravity; auto-reload, a monthly auto-payment limit, balance and transaction tracking, and invoices have also been added.

    Usage details

    Who can use it: Customers with a self-serve/online Cloud Billing account and an active, non-trial Gemini Enterprise subscription; Standard, Plus, and eligible Standard Emerging Market and Pay-as-you-go editions. The feature is rolling out gradually; reseller-managed, free trial, and free-tier accounts are not eligible.

    How to use it: In Gemini Enterprise, go to `Plans & billing` > `Credit balance` > `Add credits` to purchase credits and set up auto-reload if needed. For over-quota usage, enable the `Pay-as-you-go usage above quota` toggle under `Usage & spend` > `Usage`; also set a monthly spending limit.

    Tip: Set monthly auto-payment and project spending limits before enabling auto-reload; account for the possibility that production traffic may stop when the credit balance reaches zero.

  • ModelGemini API ve Gemini Enterprise Agent Platform

    Gemini Omni 1.2 Flash reaches GA in API and Agent Platform

    Gemini Omni 1.2 Flash is generally available in Gemini API and Google Cloud Gemini Enterprise Agent Platform. It offers video generation/editing up to 15 seconds, improved character and object consistency, multi-turn editing, and support for up to six references. Agent Platform also supports standalone video upscaling, repeatable generation with `seed`, and adding `labels` to requests; paid-tier rate limits in the API have doubled, and video output prices have fallen by 30 percent.

    Usage details

    Who can use it: GA in Gemini API, Google AI Studio, and Gemini Enterprise Agent Platform; paid tiers are supported in the API, while Cloud supports the global region with Standard PayGo and Provisioned Throughput options.

    How to use it: Select the `gemini-omni-1.2-flash` model in AI Studio, or call the Interactions API with the same model ID in Cloud Agent Platform Studio. Set the video duration/resolution; use `task: "upscale"` for upscaling in Cloud and `seed` for repeatable output.

    Tip: Generating a quick 360p/720p draft first and upscaling the best result to 1080p or 4K reduces the cost of re-generation.

  • FeatureCodeMender

    CodeMender v0.18.0 adds verification limits and IAM guidance

    CodeMender v0.18.0 adds exploratory tool-turn limits to prevent unlimited command loops in `cm verify` sessions. The update also adds a release notes link to CLI update prompts, `roles/aiplatform.agentsUser` IAM guidance for service account and CI/CD errors, and improvements to sessions, shells, input, and error reporting.

    Usage details

    Who can use it: CodeMender on Gemini Enterprise Agent Platform is available as a Public Preview to a limited group of customers.

    How to use it: Install the latest CodeMender CLI and run `cm verify`; for IAM or CI/CD setup errors, apply the `roles/aiplatform.agentsUser` binding commands suggested by the CLI and check the release notes link in the update prompt.

    Tip: Use the new default guardrails to keep verification loops limited; test in an isolated environment before disabling human approval or sandbox safety.

    ✓ Verified by official sourceModels: Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.8 Flash CyberSource: Google Cloud sürüm notları – CodeMender v0.18.0Source: CodeMender genel bakış ve CLI kullanımı
  • APIGemini API

    Function responses can return audio, video, and Office files to the model

    Tool function responses can now carry audio, video, Office/OpenDocument files, and Files API URIs, in addition to images and PDFs. The GenerateContent API supports `fileData`, `displayName`, and `mediaProcessing` for video; the Interactions API supports `document`, `audio`, and `video` blocks in `function_result`.

    Usage details

    Who can use it: Developers using Gemini API GenerateContent and Interactions API flows; the relevant version and processing options apply to video blocks.

    How to use it: In a GenerateContent function response, include the appropriate `inlineData` or `fileData` field; in the Interactions API, add a `document`, `audio`, or `video` block to `function_result`. For large files, upload them to the Files API first and return a URI, and specify the video processing mode.

    Tip: Using URIs for large videos and multi-page documents reduces the function response body and the risk of hitting request size limits.

  • APIGemini API

    Gemini API adds support for PPT/ODP/ODS and older DOC files

    Gemini API now accepts PowerPoint 97–2003 (`PPT`), OpenDocument Presentation (`ODP`), and OpenDocument Spreadsheet (`ODS`) files as input. Page image processing has been added for Word 97–2003 (`DOC`) files; `DOC`, `PPT`, and `ODP` are converted to page images, while data and formulas on the first page of `ODS` are passed to the model.

    Usage details

    Who can use it: Developers using Gemini API; the new file types are available in GenerateContent and related document processing flows.

    How to use it: Send the file in the request using a Files API URI or directly as document data, and specify the appropriate MIME type; account for page image processing for `DOC`, `PPT`, and `ODP`, and first-page data/formula processing for `ODS`.

    Tip: For documents where layout is critical, also send the same content as a PDF for comparison to verify the visual layout.

  • APIGemini Enterprise Agent Platform

    Agent Registry region mapping clarified for Gemini Enterprise apps

    Google recommends mapping each Agent Gateway in Gemini Enterprise apps to a single regional Agent Registry instance in the same project and region. Use the `us-central1` registry for `global` and `us` apps, and `europe-west1` for `eu` apps.

    Usage details

    Who can use it: Administrators managing Agent Gateway and Gemini Enterprise apps; applies to `global`, `us`, and `eu` app locations.

    How to use it: Place the Agent Gateway in the same project and region as the Gemini Enterprise app, connect a single regional Agent Registry to the gateway, and register agents, endpoints, and MCP servers in that registry; use `us-central1` for `global`/`us` and `europe-west1` for `eu`.

    Tip: Keeping agents and their targets in the registry connected to the gateway prevents resource-based IAM and governance rules from unexpectedly being bypassed.

  • AvailabilityGemini Enterprise

    Gemini Enterprise reaches allowlist GA in Brazil’s `br` region

    The Gemini Enterprise app is generally available in Brazil’s (`br`) in-country region. The region keeps data at rest and machine learning processing within the region; access is provided through an allowlist.

    Usage details

    Who can use it: Gemini Enterprise customers in Brazil; GA requires an allowlist, and customers must contact their Google account team for access.

    How to use it: When setting up a Gemini Enterprise app in Google Cloud, select the `br` in-country region and request allowlisting from your Google account team; then verify your data residency settings.

    Tip: If legal data residency is not required, compare whether the global location offers newer models and features; if it is required, make the `br` choice early in your app and data store design.

  • RetiredGemini API

    Gemini Omni 1.1 Flash requests will be routed to Omni 1.2 on December 3

    Google announced that, starting December 3, 2026, it will automatically route Gemini API requests for `gemini-omni-1.1-flash` to the `gemini-omni-1.2-flash` model. Requests that continue using the old ID will be billed at Omni 1.2 Flash rates; this is preannounced automatic routing, not an immediate shutdown.

    Usage details

    Who can use it: Projects using the `gemini-omni-1.1-flash` model ID in Gemini API; the change takes effect on 2026-12-03.

    How to use it: Change the model ID to `gemini-omni-1.2-flash` now, and test the outputs and cost; if you temporarily keep using the old ID, account for 1.2 rates in your budget after December 3.

    Tip: Switching to the fixed model ID before automatic routing begins lets you evaluate changes to output format and pricing in a controlled way.

  • Plugin / IntegrationGoogle Vids

    Gemini Omni 1.2 Flash brings 15-second clips to Google Vids

    Gemini Omni 1.2 Flash is now available in Google Vids; users can set clip durations between 3 and 15 seconds. The update improves character/product consistency in reference images, as well as speaker variety and audio realism.

    Usage details

    Who can use it: Business Starter/Standard/Plus, Enterprise Starter/Standard/Plus, Education Plus, personal Google accounts, Google AI Pro/Ultra, Enterprise Essentials, Nonprofits, Individual, and related education/AI Expanded Access add-ons; the full rollout began on November 5.

    How to use it: Open Google Vids, select `Create AI video` or `AI video` in an existing video, add prompts and references in the `Create` panel, and set the duration slider between 3 and 15 seconds.

    Tip: For branded product videos, keep the same reference images and test camera movement with short variations first.

  • AvailabilityGemini app

    15-second video generation with Omni 1.2 rolls out to Ultra in the Gemini app

    Google said that 15-second video generation with Gemini Omni 1.2 Flash is also rolling out to the Gemini app, with a phased rollout to Google AI Ultra subscribers. This is a new use case in the consumer Gemini app, in addition to API and Vids access for the model.

    Usage details

    Who can use it: Phased rollout; Google AI Ultra subscribers are the target. Google did not specify a full list of regions for this app rollout in the announcement.

    How to use it: Submit a video generation prompt in the Gemini app; when video generation and the Omni 1.2 option appear in your account, use the 15-second duration and reference images, if available.

    Tip: For a more consistent first result, provide a short, clear action and a consistent reference image rather than a complex scene all at once.

    ✓ Verified by official sourceModels: Gemini Omni 1.2 FlashSource: Google duyurusu: Gemini Omni 1.2 Flash ile daha uzun klipler
  • FeatureGoogle Forms

    Actionable insights with Gemini arrive in Google Forms

    Actionable Insights in the Responses tab of Google Forms uses Gemini to analyze responses to all question types and extract trends, correlations across questions, key findings, and recommended next steps. It writes and runs Python code on actual response data to produce counts, percentages, and comparisons.

    Usage details

    Who can use it: Business Standard/Plus, Enterprise Standard/Plus, Education Plus, Google AI/Pro/Ultra, and Google AI Pro for Education with Teaching and Learning; the Rapid Release rollout began on November 2, and Scheduled Release will begin on November 19.

    How to use it: Open a form in Google Forms, go to the `Responses` tab, and click `Generate insights` in the `Actionable Insights` section; review the calculations and recommendations under `Analysis steps`.

    Tip: Check the sample size and underlying responses before acting on recommendations; do not interpret automated correlations as causation.

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 · 7

  • APIClaude Enterprise Admin API

    Signing a member out of all Claude sessions with the Admin API enters beta

    Claude Enterprise administrators can now sign a member out of Claude sessions on all browsers and devices with a single call, without removing them from the organization. The action ends sessions, trusted devices, and organization OAuth tokens; the member can sign in again later.

    Usage details

    Who can use it: Beta for Claude Enterprise organizations. The Admin API key requires the write:members scope; API keys, already-running cloud sessions, and scheduled tasks are not terminated by this action.

    How to use it: Send a POST https://api.anthropic.com/v1/organizations/users/{user_id}/sign_out call using the member’s ID, with the anthropic-version: 2026-10-21 and anthropic-beta: admin-user-sign-out-2026-10-08 headers; no body is required. Alternatively, use the Sign out everywhere action on the Members page in organization settings.

    Tip: Prepare a write:members key in advance for security incidents; if you need to permanently block the person, also disable their account in the identity provider or remove the member.

  • FeatureClaude Enterprise

    Security Center beta released for Claude Enterprise

    Security Center compares a Claude Enterprise organization’s chat, Cowork, and Claude Code settings against Anthropic’s recommendations on a single page. It shows control status, new or recently changed settings, manual attestations, and a downloadable report for auditing.

    Usage details

    Who can use it: Beta for Claude Enterprise organizations. Only the Owner and Primary Owner roles can open the page; it is read-only and is not a certification or compliance guarantee.

    How to use it: In Claude, open the account menu and go to Organization settings > Security posture. Review the Passed, Failed, Not attested, and Shown for information statuses; use the relevant setting link for supported controls and Download report to get the report.

    Tip: Address Failed and Not attested controls first in a pilot organization; Security Center makes recommendations but does not change settings itself.

  • SettingClaude Console

    Beta IP allowlist opens for Claude organizations

    Organization administrators can now restrict organization traffic in Claude Console to allowed public IP addresses and CIDR ranges. API requests and covered application access from addresses outside the list are denied.

    Usage details

    Who can use it: Beta for Claude Platform and Claude Enterprise organizations. IPv4/IPv6 addresses or CIDR ranges can be used; there are up to 1,000 entries per organization. Bedrock, Claude Platform on AWS, and some Claude deployments on Google Cloud are not covered.

    How to use it: In Claude Platform, go to Console > Identity and access; in Claude Enterprise, go to claude.ai > Organization settings > Organization and access. Enable the IP allowlist and use Add entry or Add IP range to add the public addresses of servers, CI, VPNs, and user networks; save with Save. Changes can take up to 15 minutes to take effect.

    Tip: First list all data center, CI, VPN, and proxy egress addresses; leave the emergency access option enabled so administrators can return to the settings page from anywhere.

  • FeatureClaude

    Personal Finance becomes available in Claude as a beta

    Personal Finance shows balances and transactions from connected bank, card, credit, and investment accounts in one place, letting you ask Claude money-related questions in natural language. Access is read-only; Claude cannot move, manage, or invest money.

    Usage details

    Who can use it: Rolling beta in the US on Pro, Max, and Ultra plans. Available on web Claude, as well as Claude Desktop, Claude for iOS, and Claude for Android; not available on Free, Team, or Enterprise plans.

    How to use it: Open Personal Finance in Claude’s sidebar; if it is not visible, expand More. Follow Get started > Continue > Link accounts with Plaid, select accounts in the Plaid window, and return to Claude. You can turn Plaid access on or off in a chat from + > Connectors.

    Tip: Compare important amounts with your bank’s app; balances may be delayed and Claude’s answers may be incorrect. Enter your bank login details in the Plaid window, not in Claude.

  • Plugin / IntegrationClaude Cowork

    EDR sensor beta arrives for Claude Cowork’s local virtual machine

    An organization’s Endpoint Detection and Response sensor can run in the isolated virtual machine on the computer used by Claude Cowork and unified Claude. Claude Desktop verifies and installs the signed sensor package each time the VM starts; sensor activity is reported to the organization’s existing security console.

    Usage details

    Who can use it: Beta on Team and Enterprise plans; for Claude Desktop on macOS and Windows. Claude Desktop connected to Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry is also supported. Only the local VM on the user’s computer is covered; Anthropic’s cloud VMs are not.

    How to use it: Using MDM or Windows Group Policy/Intune, set `coworkEDRIntegration` to the full path of the EDR provider’s signed sensor package and `coworkEDRIntegrationVendor` to the provider’s lowercase identifier. Requires Claude Desktop 4.4212.0+ on macOS, and 4.4212.0+ MSIX, Windows 10/11, and virtualization support on Windows.

    Tip: Because the sensor can see prompts, files, and Claude responses and send data to the EDR provider, first verify the vendor’s data flows and detection-only behavior.

  • FeatureClaude Enterprise

    Centralized management console enters beta for Enterprise organizations

    Centralized management lets a company manage its connected Claude Enterprise and Claude Platform organizations from a single parent organization console at admin.claude.com. It brings together SSO, SCIM, domain verification, groups, Compliance API keys, multiple identity providers, and new parent roles in one place.

    Usage details

    Who can use it: Beta on Enterprise plans. Child organizations retain separate user, role, project, and data boundaries; Claude Platform organizations purchased through the AWS or Azure marketplace cannot join a parent.

    How to use it: An administrator of a child organization starts setup at admin.claude.com and nominates candidates for parent organization owner/admin. After administrators of the other child organizations approve the nomination, enrollment is complete; because setup cannot be reversed, clarify the candidates and domain/identity-provider structure in advance.

    Tip: First check the impact of SSO, SCIM, and tenant restriction in a sandbox or small organization; a parent role does not automatically grant access to a child organization.

  • SettingClaude Desktop

    Claude Desktop access controls arrive for Enterprise administrators

    Claude Enterprise administrators can now manage common Claude Desktop settings through Organization settings; block personal and other external Claude accounts on managed computers; and restrict which working folders Claude can use. This is part of a beta access-control package aimed at reducing shadow IT.

    Usage details

    Who can use it: Beta for Claude Enterprise; for Claude Desktop users on managed computers. MDM, a macOS configuration profile, or Windows Group Policy/Intune can also be used for folder, organization account, local MCP, and extension restrictions.

    How to use it: Turn on the Claude Desktop controls in Organization settings in Claude. To enforce them through device policy, use the com.anthropic.claudefordesktop profile on macOS or the HKLM/HKCU Claude policy path on Windows; configure allowedWorkspaceFolders to restrict folders and forceLoginOrgUUID to require organization sign-in.

    Tip: Before blocking external accounts, test SSO, local MCP, and extension requirements on pilot devices; otherwise, legitimate developer workflows may be interrupted.

OpenAI · ChatGPT

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

New in this edition · 10

  • RetiredOpenAI API

    API shutdown for gpt-5.5-cyber-preview announced for February 1, 2027

    OpenAI marked `gpt-5.5-cyber-preview` for deprecation as of November 5, 2026. The model will be removed from the API on February 1, 2027; the official recommendation is to switch to the most capable cybersecurity model available to the account.

    Usage details

    Who can use it: All projects using `gpt-5.5-cyber-preview` in the API are affected; removal date is 2027-02-01.

    How to use it: Find uses of `gpt-5.5-cyber-preview` in code, configuration, and model logs. If you have access, evaluate `gpt-6-cyber` or the most capable cybersecurity model available to your account in staging, then complete the migration before February 1, 2027.

    Tip: Run comparative security and cost tests with both models before the shutdown date; define a fallback model subject to access approval.

    ✓ Verified by official sourceModels: gpt-5.5-cyber-previewSource: OpenAI API deprecation listesi
  • ModelOpenAI API ve Codex

    GPT-6 Cyber released for authorized cybersecurity research

    OpenAI introduced GPT-6 Cyber, a model that can find and validate vulnerabilities and is focused on staying within the scope set by the researcher. The model reported a 51% relative improvement over GPT-5.6 Cyber on ExploitGym, and its Preparedness Framework classification for cybersecurity was listed as Critical.

    Usage details

    Who can use it: Initially available through the API and Codex to select, approved Daybreak Red customers; a phased expansion to all approved Daybreak Red customers is planned. GPT-6-Cyber also requires model approval, project access, an IP allowlist, and applicable organization security requirements.

    How to use it: Complete Daybreak Red and GPT-6-Cyber model approval with your account team; enable access for the project and configure the IP allowlist, SSO/MFA, and Private Safety Processing if required. In a Responses API request, send `model: "gpt-6-cyber"`, `access_programs.cyber: "daybreak_red"`, and a stable `safety_identifier`.

    Tip: Run initial tests in an isolated test environment with explicit rules and scope boundaries; use human approval and action review for high-risk actions.

  • APIOpenAI Go SDK

    Go SDK added data residency constants for AU, IN, and JP

    OpenAI Go SDK 3.95.0 added data residency options for selecting the Australia, India, and Japan endpoints. The new constants can be used with `WithDataResidency`.

    Usage details

    Who can use it: Go projects using the OpenAI API; regional processing and organization/project eligibility requirements apply.

    How to use it: Run `go get github.com/openai/openai-go/v3@v3.95.0`. Configure the client with `WithDataResidency(openai.DataResidencyAU)` or the corresponding `DataResidencyIN`/`DataResidencyJP` option.

    Tip: Before deploying a regional endpoint change, verify that the application uses the same policy across all OpenAI client instances.

  • APIOpenAI Node SDK

    Node SDK expands agent errors, image compression, and audio options

    Node SDK 7.51.0 recognizes the `max_output_tokens` code in agent session turn errors, and adds image-context compression options for Responses requests and noise-reduction strength options for Realtime input audio. The changes apply to stable and beta Responses types and helper flows.

    Usage details

    Who can use it: Agents, Responses, and Realtime API integrations using OpenAI Node SDK 7.51.0.

    How to use it: Update with `npm install openai@7.51.0`. Set the image compression option in the Responses request object and the noise-reduction `strength` value in Realtime audio settings; handle `max_output_tokens` in the agent error union.

    Tip: Compare WebSocket, HTTP, and parse/stream helpers using the same test inputs to verify that the new fields are passed through on every transport path.

  • APIOpenAI Node SDK

    Node SDK adds support for data residency in Australia, India, and Japan

    OpenAI Node SDK 7.51.0 added Australia, India, and Japan values to the `dataResidency` client option and updated the documentation for the seven supported regions. This allows regional endpoint selection from the Node client for eligible API projects.

    Usage details

    Who can use it: Node.js projects using the OpenAI API; regional processing and organization/project eligibility requirements apply.

    How to use it: Update the package with `npm install openai@7.51.0`. Initialize the client with `dataResidency: "au"`, `"in"`, or `"jp"` and check your project's data residency authorization.

    Tip: Roll out gradually, verifying that region-based endpoint selection is reflected in observability and error logs.

  • APIOpenAI Python SDK

    Python SDK expands agent errors, image compression, and audio options

    Python SDK 3.46.0 recognizes the `max_output_tokens` code in agent session turn errors and supports image-context compression options for Responses requests and noise-reduction strength options for Realtime input audio. The `responses.parse()` and `responses.stream()` helpers also pass the image option into the request body.

    Usage details

    Who can use it: Agents, Responses, and Realtime API integrations using OpenAI Python SDK 3.46.0.

    How to use it: Run `pip install --upgrade openai==3.46.0`. Set the image-compression option in Responses requests and the input-audio noise-reduction `strength` value in Realtime requests; handle agent turn errors to include the `max_output_tokens` case.

    Tip: Test streaming and parse flows separately; verify that existing default behavior is preserved when image and audio fields are omitted.

  • APIOpenAI Python SDK

    Python SDK added support for data residency in Australia, India, and Japan

    OpenAI Python SDK 3.46.0 added Australia, India, and Japan as values for the `data_residency` option when creating or copying a client. This lets eligible projects select an API endpoint for one of these regions through the SDK.

    Usage details

    Who can use it: Python projects using the OpenAI API; regional processing, project eligibility, and organization controls apply.

    How to use it: Run `pip install --upgrade openai==3.46.0`. Create or copy the client with `data_residency="au"`, `"in"`, or `"jp"`, and verify your project's access to the relevant region.

    Tip: Before using a region selection in production, verify data retention, latency, endpoint, and backup assumptions in a staging environment.

  • APIOpenAI Go SDK

    Go SDK added a model option and model constants to Responses types

    Go SDK 3.96.0 added the `Model` option to the generated types for Responses requests and updated model constants and package aliases. Existing free-form model values remain supported.

    Usage details

    Who can use it: Responses API applications using OpenAI Go SDK 3.96.0.

    How to use it: Run `go get github.com/openai/openai-go/v3@v3.96.0` and set the model field with `responses.ResponseNewParams{Model: "gpt-6-astra", ...}`.

    Tip: Centralize model selection in a single configuration layer; run build and request-serialization tests together after upgrading the SDK.

  • APIOpenAI Node SDK

    Node SDK adds a model option to Responses types

    Node SDK 7.52.0 added the `model` option to the generated types for Responses requests. Shared model unions were expanded, while compatibility with existing model values as plain strings was preserved.

    Usage details

    Who can use it: Responses API applications using OpenAI Node SDK 7.52.0.

    How to use it: Run `npm install openai@7.52.0` and specify the model in a call such as `client.responses.create({ model: "gpt-6-astra", input: ... })`.

    Tip: After updating the types, add a CI check to catch any model constants you use at build time.

  • APIOpenAI Python SDK

    Python SDK added a model option to Responses types

    Python SDK 3.47.0 added the `model` option to the generated types for Responses requests. Backward compatibility was maintained by preserving shared model unions and free-text model values.

    Usage details

    Who can use it: Responses API applications using OpenAI Python SDK 3.47.0.

    How to use it: Run `pip install --upgrade openai==3.47.0` and specify the model directly in a call such as `client.responses.create(model="gpt-6-astra", input=...)`.

    Tip: Keep the model name configurable rather than hard-coding it in production; run regression tests for existing free-text model values during migration.

Meta · Meta AI

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

New in this edition · 8

  • CapabilityMuse

    Real-time voice conversations and hundreds of voice options come to Muse

    Meta is rolling out real-time voice conversations in the Muse app on mobile, web, and Mac, starting in the US. Users can interrupt while speaking; Muse can continue the conversation and work on a task, and users can choose from hundreds of different voices.

    Usage details

    Who can use it: Starting in the US; mobile, web, and Mac versions of the Muse app. Voice conversations are coming soon to AI Glasses and Muse Charm.

    How to use it: Open Muse on mobile, web, or Mac; tap the phone icon in the top right to start a voice conversation and choose the voice you want.

    Tip: Handle tasks such as planning a trip, brainstorming, or interview practice in a continuous voice-based flow; interrupt Muse’s response and ask a follow-up question when needed.

  • AvailabilityWhatsApp Events

    WhatsApp Events invitations are now open to people without WhatsApp accounts

    Event invitations can now reach people without WhatsApp accounts through a shareable link. Guests can view event details and RSVP without downloading the app; their phone numbers aren’t shown to guests they don’t know, and the address is revealed to guests without accounts after they RSVP.

    Usage details

    Who can use it: The updated Events feature is rolling out to everyone over the following weeks; guests without WhatsApp accounts can join via a link.

    How to use it: Create an Event in WhatsApp, add a cover photo and details, then send the event link to people outside WhatsApp; guests open the link and RSVP.

    Tip: Require an RSVP before sharing the address; this prevents guests without accounts from seeing the location too early.

  • FeatureWhatsApp Events

    WhatsApp Events Home brings event and RSVP management together in one place

    WhatsApp is introducing a new Events Home that brings event creation, customized invitations, and RSVP management together in one place. Events Home also lets users manage past and upcoming events, custom reminders, and the plus-one option.

    Usage details

    Who can use it: The updated Events experience is rolling out gradually to all users over the following weeks.

    How to use it: In WhatsApp, go to Settings > Events; create an event, add a cover photo if you want, share the invitation, and manage RSVPs from the same screen.

    Tip: Add a custom reminder to help prevent attendees from forgetting; enable the plus-one option only for events with sufficient capacity.

  • FeatureMuse

    Muse avatars respond to conversations with facial expressions and gestures

    Muse’s avatar displays facial expressions and gestures that match the tone of the conversation while listening and speaking. Users can choose audio only, an avatar alongside the chat, or a full-screen face-to-face view.

    Usage details

    Who can use it: Rolling out in the US first on the mobile, web, and Mac versions of the Muse app; the same announcement says voice conversations are coming soon to AI Glasses and Muse Charm.

    How to use it: Open a call using the phone icon in the top right of Muse; in the call view, choose voice only, avatar alongside, or full screen mode.

    Tip: Use voice-only for tasks that require focus; for rehearsals or conversations, turn on the avatar and full-screen view for visual feedback.

  • AvailabilityMeta Quest

    PAYDAY: Aces High launches on Meta Quest 3 and 3S

    PAYDAY: Aces High has launched for Meta Quest 3 and 3S. The game costs 24,99 ABD doları on the Meta Store; Meta Horizon+ subscribers can download it from the subscription catalog on launch day.

    Usage details

    Who can use it: Meta Quest 3 and Meta Quest 3S; access through purchase or a Meta Horizon+ subscription.

    How to use it: Open the PAYDAY: Aces High page in the Meta Quest Store, then purchase and install it; if you’re a Meta Horizon+ subscriber, download the game from the subscription catalog.

    Tip: The game supports co-op for up to four players; plan your team’s stealth, weapons, and role assignments before starting a mission.

  • RetiredGraph API

    Retirement dates announced for Graph API v21.0 and v22.0

    Meta announced that Graph API v21.0 will be retired and removed on 21 January 2027, and v22.0 on 20 May 2027. Applications relying on these versions need to move to the latest Graph API version.

    Usage details

    Who can use it: All applications and integrations using Graph API v21.0 or v22.0.

    How to use it: Identify the Graph API version used in your code and configuration; follow the upgrade guide for the latest version and test calls in a staging environment.

    Tip: Don’t wait until the deadline; add the version number to CI checks and regularly scan for use of old endpoints and fields.

  • RetiredFacebook Platform SDK

    2027 retirement schedule for Facebook Platform SDK v21.0 and v22.0

    Meta Business SDK v27 notes in the Graph API v27 announcement say Facebook Platform SDK v21.0 will be retired in January 2027 and v22.0 in May 2027. Developers who pin their SDK version need to plan a move to supported versions.

    Usage details

    Who can use it: Developer applications using Facebook Platform SDK v21.0 or v22.0.

    How to use it: Inventory the Platform SDK version you use, upgrade to a supported version, and test login, permissions, webhooks, and Graph API calls.

    Tip: Perform a controlled upgrade instead of leaving dependency versions to update automatically; add the January and May 2027 deadlines to your CI schedule.

  • APIPyrefly

    Pyrefly v1.4.1 makes the coverage command stable and fast

    Meta’s open-source Python type checker Pyrefly v1.4.1 fixes inconsistent results, panics when an output stream closes early, and incorrect typed/untyped reports for `pyrefly coverage`. The release also recognizes Cython pure-Python declarations and speeds up coverage runs in large projects.

    Usage details

    Who can use it: Python developers using Pyrefly; `pyrefly coverage check` can now also be used as a pre-commit hook.

    How to use it: Run `pip install --upgrade pyrefly==1.4.1`; then use `pyrefly coverage check` or `pyrefly-coverage-check` in pre-commit.

    Tip: After upgrading, run coverage several times in CI to verify deterministic output and gradually clean up type errors.

    ✓ Verified by official sourceSource: GitHub — Pyrefly v1.4.1 Release

Moonshot AI · Kimi

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

New in this edition · 32

  • CapabilityKimi Code CLI

    Requirement to reread a file before Edit removed

    Kimi Code removed the restrictions that required using the Edit tool and rereading a file when making consecutive edits to the same file.

    Usage details

    Who can use it: Kimi Code 2.6.0 canary and later versions.

    How to use it: In the current CLI, after adding a file to context once, run Edit directly on the same file.

    Tip: Although the restriction has been removed, check the file’s current state and the diff before Edit when making critical changes.

  • CapabilityKimi Code CLI

    Session status and Git data in remote environments come from the connected environment

    When using a remote environment, Kimi Code now routes session status and Git data through the connected work environment.

    Usage details

    Who can use it: Kimi Code 2.6.0 canary users using Remote environment.

    How to use it: Sign in to a remote environment with version 2.6.0 canary; status and Git information will be shown using data from the connected environment.

    Tip: If the status in a remote repository looks inconsistent, first verify the connected environment’s working directory and Git access.

  • CapabilityKimi Code

    Image/video previews fixed for live and reloaded tool results

    Tool results now show image and video previews that were missing during streaming or after reloading a session.

    Usage details

    Who can use it: Kimi Code sessions containing image/video tool results; 2.5.1 and later.

    How to use it: Update Kimi Code to 2.5.1 or later; generate a tool result or reopen an old session.

    Tip: If the preview does not appear, reload the session and reopen the relevant tool card.

  • APIKimi Code CLI

    Repeated HTTP 400 errors in dynamic tool schemas fixed

    Repeated HTTP 400 errors caused by an empty system message in sessions using dynamic tool schema have been fixed.

    Usage details

    Who can use it: Kimi Code sessions using dynamic tool schema; Kimi Code CLI 2.5.1 and later.

    How to use it: Update Kimi Code CLI to 2.5.1 or later; reauthorize the dynamic tools and run the session again.

    Tip: If the 400 error persists, log the system messages sent in the request together with the dynamic tool schema.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.5.1 GitHub sürümü
  • CapabilityKimi Work

    Goal mode progress is now directly above the input box

    Goal mode task progress is now shown directly above the message input area while a task is running.

    Usage details

    Who can use it: Kimi Work 3.2.25 Goal mode users.

    How to use it: Start a Goal mode task in Kimi Work and track its progress indicator above the Composer.

    Tip: For long goals, use the progress bar alongside interim checkpoints to see whether the task is stalled.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • Plugin / IntegrationKimi Work

    Kimi Work detects unauthorized plugins during task execution and guides users

    Unauthorized plugins are automatically detected while a task is running, and users are shown a pop-up with authorization guidance.

    Usage details

    Who can use it: Kimi Work 3.2.25 users who use plugins.

    How to use it: Start a task that requires a plugin; when the unauthorized plugin notification appears, complete the authorization steps in the pop-up.

    Tip: Granting plugin permissions before starting a task reduces interruptions during execution.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • FeatureKimi Work

    Kimi Work adds the Stop and Archive action

    The new Stop and Archive action stops the ongoing task before archiving, then archives the session.

    Usage details

    Who can use it: Kimi Work 3.2.25; users with running tasks.

    How to use it: Open the menu for a running task or session, select Stop and Archive, and confirm.

    Tip: If you want to save undelivered output before archiving, check the task’s latest status first.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • OtherKimi Work

    Compatibility issues with local artifacts, ZIP files, and path authorization on Windows fixed

    Various compatibility issues involving local artifact links, ZIP files, and path authorization on Windows have been fixed.

    Usage details

    Who can use it: Kimi Work 3.2.25 on Windows.

    How to use it: Update Kimi Work to 3.2.25 and retry the local artifact link, ZIP file, or authorized path workflow.

    Tip: Reopen the artifact link after updating Windows path separators and access permissions.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • FeatureKimi Code Web

    The session info panel no longer covers the chat in narrow Web layouts

    Fixed an issue where the session info panel automatically floated over and obscured chat messages in narrow layouts.

    Usage details

    Who can use it: Narrow window or panel layouts in Kimi Code Web; 2.5.1 and later.

    How to use it: Update the Web interface to 2.5.1 or later, then open the session info panel at a narrow window width.

    Tip: To read details on a narrow screen, temporarily widen the window instead of repeatedly opening and closing the panel.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.5.1 GitHub sürümü
  • FeatureKimi Code Web

    Stale empty implicit workspaces are removed from the list

    Kimi Code Web now hides unused, empty implicit workspace entries from the workspace list.

    Usage details

    Who can use it: Kimi Code Web workspace list; 2.5.1 and later.

    How to use it: Update Web to 2.5.1 or later, then reopen the workspace list.

    Tip: Keep the list tidy by closing temporary sessions you no longer use and refreshing the list.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.5.1 GitHub sürümü
  • CapabilityKimi Code CLI

    Fullscreen TUI groups tool calls and thinking in transcripts

    Fullscreen TUI provides a cleaner view by collapsing tool calls and thinking content into transcript groups.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 canary users using Fullscreen TUI.

    How to use it: Update Kimi Code to version 2.6.0 canary and open TUI mode as fullscreen; expand groups from the transcript.

    Tip: In long sessions, expand only the tool group you want to inspect to make it easier to navigate the context.

  • OtherKimi Work

    Kimi Work installs and launches faster

    Version 3.2.25 improved installation and app startup speed.

    Usage details

    Who can use it: Kimi Work 3.2.25; supported desktop platforms.

    How to use it: Update Kimi Work to 3.2.25 or perform a clean install; the change applies automatically.

    Tip: For major updates, verify that the new version launches before deleting old installation files.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • FeatureKimi Work

    Kimi Work moved chat history navigation to the left

    Chat history navigation controls are now on the left side of the interface.

    Usage details

    Who can use it: Kimi Work 3.2.25 desktop interface.

    How to use it: Open a chat and use the history navigation controls on the left in a long conversation.

    Tip: In long conversations, use navigation with search to quickly return to the section you’re looking for.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • CapabilityKimi Work

    The message queue interface no longer interrupts the reading flow

    Kimi Work improved the interface behavior that caused the message queue to interrupt the flow while reading a conversation.

    Usage details

    Who can use it: Kimi Work 3.2.25; users who use the message queue while a response is in progress.

    How to use it: Send a new message while the agent is responding; you’ll see that the queue interface no longer blocks you from reading the conversation.

    Tip: Review the queue and remove unnecessary prompts before sending multiple messages.

    ✓ Verified by official sourceSource: Kimi Work 3.2.25 sürüm notları
  • OtherKimi Code CLI

    Wrapped Grep and Glob lines preserve correct indentation

    Indentation has been fixed for Grep and Glob summary lines that wrap because they do not fit on a single line in the TUI.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 canary TUI.

    How to use it: Update the CLI to 2.6.0 canary and view long Grep/Glob results in the TUI.

    Tip: For very long results, narrow the search to reduce line wrapping and visual clutter.

  • FeatureKimi Code Web

    Sessions from deleted workspaces are hidden from lists

    Sessions belonging to deleted workspaces no longer appear in session lists.

    Usage details

    Who can use it: Kimi Code Web/workspace session lists; 2.5.1 and later.

    How to use it: After deleting a workspace, update Kimi Code to 2.5.1 or later and refresh the session list.

    Tip: If you may need a session later, archive and export it before deleting the workspace.

  • CapabilityKimi Code Web

    Tool summaries no longer keep expanding and collapsing while a response is streaming

    Fixed an issue where tool call summaries repeatedly expanded and collapsed while a response was streaming in the Web interface.

    Usage details

    Who can use it: Kimi Code Web transcript; 2.5.1 and later.

    How to use it: Update Web to 2.5.1 or later and watch the tool call cards during streaming.

    Tip: For long outputs, expand only the tool cards you need to keep the transcript more readable.

  • FeatureKimi Code Web

    Automatic title generation works again for new Web sessions

    Fixed an issue where new Web sessions did not get an automatic title after the first turn.

    Usage details

    Who can use it: Kimi Code Web users; Kimi Code CLI 2.5.1 and later.

    How to use it: Update the CLI/Web component to 2.5.1 or later and send the first turn in a new session; a title will be generated automatically.

    Tip: If you need a custom name, rename the session manually after the automatic title is created.

  • CapabilityKimi Code Web

    Local images with spaces in their paths now display on the Web

    An issue preventing local images with spaces in their file paths from appearing in the Web interface has been fixed.

    Usage details

    Who can use it: Sessions using local images in Kimi Code Web; 2.5.1 and later.

    How to use it: Update Kimi Code Web to 2.5.1 or later and add the local image whose path contains spaces to the session.

    Tip: After adding the image, wait for the preview to load; if it does not, select the path again with quotation marks.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.5.1 GitHub sürümü
  • CapabilityKimi Code CLI

    @ file mentions now include files in additional working directories

    In multi-directory workspaces, file mentions using `@` now include files in additional working directories.

    Usage details

    Who can use it: Kimi Code projects configured with multiple workspace directories; 2.5.0 and later.

    How to use it: Add additional directories to the project, type `@` in Composer, and select a file from an additional working directory.

    Tip: For files with the same name, choose the correct workspace path to avoid using the wrong file as context.

    ✓ Verified by official sourceSource: Kimi Code 2.5.0 GitHub sürümü
  • Plugin / IntegrationKimi Code ACP

    File tools restored in legacy ACP sessions

    Kimi Code’s ACP server fixed an issue that caused local file tools to disappear in legacy ACP sessions.

    Usage details

    Who can use it: Users who use Kimi Code with ACP clients; 2.5.0 canary and later versions.

    How to use it: Connect your ACP client to Kimi Code 2.5.0 or later, then call the file tools again in the existing session.

    Tip: If the ACP client uses the legacy session startup flow, update the client as well.

  • APIKimi Code Server API

    A minimum wire reader version is now required for Kimi Code remote communication

    Kimi Code added a minimum reader version check to block incompatible older wire reader clients and clarified upgrade guidance.

    Usage details

    Who can use it: Clients using a wire protocol, such as Kimi Code Server API/ACP; 2.5.0 canary and later versions.

    How to use it: Upgrade the server and wire reader components together to the latest version; follow the upgrade guidance shown when a compatibility error occurs.

    Tip: Pin server and client versions in deployments and upgrade them in the same update package.

  • FeatureKimi Code CLI

    Side Chat first-response and history-ordering bugs fixed

    Fixed an issue where the first Side Chat response could be mixed up with or lost alongside the main conversation's response. Also fixed Side Chat history disappearing or being reordered when the server session was loaded from disk.

    Usage details

    Who can use it: Web/Desktop sessions using Side Chat; Kimi Code 2.5.0 and later.

    How to use it: Update Kimi Code and open and use Side Chat in a session as usual.

    Tip: When a session is restored, check responses in the relevant Side Chat panel rather than the main chat.

    ✓ Verified by official sourceSource: Kimi Code 2.5.0 GitHub sürümü
  • AvailabilityKimi Code CLI

    Remote environments now support Windows targets

    Windows support has been added for Kimi Code remote environments created using SSH, Docker, and custom commands.

    Usage details

    Who can use it: Users running Kimi Code on Windows and connecting to a remote environment; 2.5.0 canary and later versions.

    How to use it: Use `--environment <id>` when creating a session; select a remote connection configuration suitable for the environment's Windows target.

    Tip: Explicitly specify Windows shell and path formats for commands running in the remote environment.

  • CapabilityKimi Code Desktop

    Local file tools fixed in existing Desktop sessions after an upgrade

    An issue that prevented local file tools from working in existing sessions after upgrading Kimi Code Desktop has been fixed.

    Usage details

    Who can use it: Existing sessions using Kimi Code Desktop; Kimi Code 2.5.0 and later.

    How to use it: Update Desktop with components 2.5.0 or later; use the file tools again without closing the existing session.

    Tip: If the issue persists after upgrading, reload the session first, then run the file tool again.

    ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 sürüm notları
  • OtherKimi Code CLI

    Three bugs in search, completion, and directory warnings fixed

    Kimi Code fixed a search bug with dot-only filenames, completion exceeding its timeout, and incorrect directory warnings.

    Usage details

    Who can use it: Kimi Code CLI 2.5.0 and later.

    How to use it: Update the CLI to 2.5.0 or later; use file search and path completion as usual.

    Tip: Use the updated CLI when searching for filenames that begin with a dot or consist only of dots.

  • APIKimi Code CLI

    Unchanged custom registry providers are no longer shown as updated

    When the model catalog is refreshed, a custom registry provider is no longer incorrectly reported as updated if nothing has actually changed.

    Usage details

    Who can use it: Kimi Code CLI users with custom `api.json` model registries; 2.5.0 and later.

    How to use it: Refresh the custom provider catalog with `/provider refresh <providerId>`; if nothing has changed, you’ll no longer receive a false update notification.

    Tip: After refreshing, review the local configuration only for actual model changes.

  • CapabilityKimi Code CLI

    UTF-8 BOM is preserved when editing files

    Kimi Code no longer loses an existing UTF-8 byte-order mark when editing files.

    Usage details

    Who can use it: Kimi Code users editing files that contain a UTF-8 BOM; 2.5.0 canary and later versions.

    How to use it: Edit and save the file with Kimi Code; BOM preservation is applied automatically.

    Tip: If language tools are BOM-sensitive, verify the file's first bytes with a project check after making changes.

  • CapabilityKimi Code CLI

    Hidden media tool outputs now clearly show an omission notice

    Unshown results from media tools are now represented in the transcript by an omission notice instead of disappearing silently.

    Usage details

    Who can use it: Kimi Code sessions using tools that produce media output; 2.5.0 canary and later versions.

    How to use it: Run a media tool with Kimi Code 2.5.0 or later; distinguish unshown results by reading the notice in the transcript.

    Tip: If you need the missing media content, rerun the tool and ask it to save the output as a file.

  • OtherKimi Code CLI

    Todo list now returns to the correct state after undo

    Fixed the todo list retaining its new state or becoming incorrect after an action was undone; the list now reverts along with the undo operation.

    Usage details

    Who can use it: Kimi Code sessions using a todo list; 2.5.0 and later.

    How to use it: Update the CLI, make a todo change, and use undo when needed to check the list.

    Tip: After undoing, verify the todo list's state before moving on to the next step.

  • CapabilityKimi Code CLI

    Long-running subagent cards remain responsive as streams grow

    Fixed long-running subagent cards becoming unresponsive as their streamed output grew.

    Usage details

    Who can use it: Kimi Code Web/Desktop/CLI surfaces using subagents; 2.5.0 and later.

    How to use it: Update Kimi Code and monitor a long-running subagent task in `/tasks` or the relevant Agent/Subagent view.

    Tip: For very long outputs, use the summary and status indicators instead of leaving the card open.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 14

  • APIHy-Nomas

    Rollback and KV snapshot/restore APIs added for Hy-Nomas sessions

    The GPU worker now provides the `/v1/rollback`, `/v1/session/state`, `/v1/session/snapshot` and `/v1/session/restore` endpoints for interactive session lifecycle management. Rollback removes the last segment; snapshot saves the KV cache, mask, noise, counters, segment metadata and PixelUMM streaming state, allowing work to resume on the same worker.

    Usage details

    Who can use it: For self-hosted Hy-Nomas web gateway/GPU worker deployments; snapshot/restore currently requires `world_size == 1`, and only trusted local files produced by the same worker should be restored.

    How to use it: With the worker ready, send `{"session_id":"..."}` in the body of `POST /v1/rollback`. Use `GET /v1/session/state` to check the state; use `POST /v1/session/snapshot` with `{"session_id":"...","path":"..."}` to save, and `POST /v1/session/restore` with the same fields to resume.

    Tip: Do not load snapshot files from untrusted sources; restore validates the file first, but the session ID and worker match are still required.

  • Plugin / IntegrationEmoReward

    Integration kit tailored to coding agents added for EmoReward

    The EmoReward repository now includes `agent/index.html`, skill workflows for porting and verification, a host adapter example, and reward contract documentation. This lets a coding agent port EmoReward into a project while preserving the existing training loop and validate reward outputs for length and finite values.

    Usage details

    Who can use it: Any local coding agent and EmoReward-based training repository; the static HTML kit requires no separate online service.

    How to use it: Tell your agent to read `agent/index.html` in the EmoReward repository, follow the linked port/verify/memory workflows, and validate the reward contract before training while preserving the existing training loop.

    Tip: Do not let the agent automatically install provider weights or global skills; first have it validate the lengths of `video_paths`, `prompts` and metadata at the host boundary.

    ✓ Verified by official sourceModels: HunyuanVideo-1.5Source: EmoReward agent entegrasyon kiti commit'i
  • APIHy-Nomas

    Hy-Nomas inference gains Tensor Parallel and torch.compile support

    Tensor Parallel support has been added to Hy-Nomas's self-hosted inference path, along with a new performance path that can be used with `torch.compile`. Official examples use `--sp-size 1 --tp-size 8` for eight ranks; in multi-machine deployments, the total process count must match `sp_size × tp_size`.

    Usage details

    Who can use it: For self-hosted multi-GPU Hy-Nomas inference or web demo deployments; limitations in the current commit include cases without TP for INT8, LoRA and some single-pass paths.

    How to use it: Set the model and VAE paths, then run `bash scripts/inference/inference_from_config.sh configs/experiment/inference.yaml --sp-size 1 --tp-size 8 ...`. For the web demo, also add `--sp-size 1 --tp-size 8` to the `start_seg_ar_demo.sh` call.

    Tip: Match `sp_size × tp_size` to the actual GPU/process count; if you use INT8 or LoRA, check this commit's TP limitations.

    ✓ Verified by official sourceModels: Hy-NomasSource: Hy-Nomas Tensor Parallel ekleyen resmi commit
  • CapabilityHy-Nomas

    PixelUMM decoding sped up in Hy-Nomas interactive previews

    PixelUMM streaming preview now reuses the pending queue instead of decoding previously decoded windows again. In the official measurement on a single GPU, preview push time in an uncompiled stream fell from 2.0–2.5 seconds to 1.0–1.5 seconds, while the output remained bit-identical.

    Usage details

    Who can use it: The Hy-Nomas interactive streaming preview path; the new behavior is enabled by default with `reuse_pending_tail=True`, and the old path can be retained with a one-line setting.

    How to use it: Update your Hy-Nomas code; `WindowStreamingDecodeSession` uses the default automatically. For a fallback comparison, set `reuse_pending_tail=False` on the session object.

    Tip: For low latency, leave the default queue-reuse behavior enabled; test visual equivalence with your own GPU and resolution profile.

    ✓ Verified by official sourceModels: Hy-NomasSource: Hy-Nomas PixelUMM streaming performans commit'i
  • FeatureHy-Nomas

    Image-to-video workflow added to the Hy-Nomas web demo

    An official Hy-Nomas commit adds an image-to-video workflow to the interactive demo. The current demo documentation explains how to upload a reference image and use Nomas-PE to generate the first five-second world, then continue in the same world with text or camera instructions.

    Usage details

    Who can use it: Hy-Nomas users running their own GPU worker/gateway deployment; the DCAE-V encoder checkpoint and Nomas-PE service are required. The demo uses a reference image only when starting a new interactive world.

    How to use it: Start the gateway with `model.vae.encoder_checkpoint` and set `PE_API_URL` and `PE_MODEL=Nomas-PE`. In the web demo, use Upload image to upload a JPEG/PNG/WebP, enter a User prompt and click Enter world; then continue with a Camera path or text instructions.

    Tip: Use the reference image to start a new world; open a new session with New world rather than changing the image of the current session.

  • AvailabilityHy-Nomas

    Hy-Nomas supports high-speed inference across four GPUs on B300

    The same official commit adds NVIDIA B300 support, B300 benchmark options and a standalone four-GPU inference script. The repository note reports measurements exceeding 24 FPS on four B300s; this is a hardware-specific benchmark excluding initial warm-up, not a general throughput guarantee.

    Usage details

    Who can use it: Four NVIDIA B300 GPUs in your own Hy-Nomas deployment, along with the Blackwell/MagiAttention, FP8 VAE and build dependencies specified in the commit.

    How to use it: In the current Hy-Nomas repository, use the `scripts/inference/inference_b300_sp4.py` entry point in an environment with four B300 GPUs; the script was added to accept external JSON input and produce MP4 output. Verify the local parameters with `--help` before running it.

    Tip: Compare FPS using the same resolution, warm-up and MP4 writing criteria; do not interpret the commit's benchmark scope as general service capacity.

    ✓ Verified by official sourceModels: Hy-NomasSource: Hy-Nomas B300 ve yüksek hızlı çıkarım resmi commit'i
  • AvailabilityHy4

    Hy4 became available across Tencent products and through OpenRouter, Hermes, and OpenCode

    Hy4 was announced as available in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant (Dayuan), and ima. API access was opened through OpenRouter, Hermes, and OpenCode; integrations with QQ, Honor of Kings, Game of Peace, WeRead, and Tencent News are still in development, according to the announcement.

    Usage details

    Who can use it: Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant (Dayuan), and ima; third-party APIs via OpenRouter, Hermes, and OpenCode. The official announcement did not specify country or plan restrictions.

    How to use it: Select Hy4 in the model picker in the Tencent product; for TokenHub, use `modelId=hy4`. On OpenRouter, the model ID is `tencent/hy4`, the endpoint is `https://openrouter.ai/api/v1/chat/completions`, and authentication uses a Bearer API key.

    Tip: Choose the Tencent integration for in-product workflows, or OpenRouter/Hermes/OpenCode if you need an OpenAI-compatible client; check the provider’s current limits and pricing before making a call.

  • Pricing / PlanTencent Cloud TokenHub

    Hy4 gets a 50% discount on Tencent Cloud TokenHub through December 31

    Tencent’s November 4 announcement offers a 50% discount on the standard price of Hy4 on Tencent Cloud TokenHub through December 31, 2026. The announcement does not provide details on eligible plans, regions, or coupon application.

    Usage details

    Who can use it: Tencent Cloud TokenHub Hy4; according to the 2026-11-04 announcement, the discount runs through 2026-12-31. Eligibility and payment terms may vary by TokenHub account.

    How to use it: Open the `modelId=hy4` model page on Tencent Cloud TokenHub and check whether the promotion is applied automatically on the payment/Token Plan screen.

    Tip: Recheck the promotion terms at checkout; the TokenHub page may also show a different, undated promotional headline, so don’t rely solely on the announcement.

    ✓ Verified by official sourceModels: Hy4Source: Tencent'in TokenHub %50 indirim duyurusu
  • FeatureMemHub

    MemHub 1.0.6 adds memory modes and multi-platform support

    MemHub 1.0.6, build 41, moved to `hy-memory==2.0.14` and announced Linux x64/arm64 Runtime and macOS arm64 App/Runtime support. Additions include Lite or Pro selection during onboarding, Ultra in Settings, 11 interface languages, validation and rollback when changing the embedding model, and more robust writing and import workflows.

    Usage details

    Who can use it: MemHub 1.0.6; Linux x64/arm64 Runtime and macOS arm64 App/Runtime. New installations offer a Lite/Pro onboarding choice; Ultra is an advanced Settings option.

    How to use it: Open MemHub 1.0.6 and select Lite or Pro during onboarding. For Ultra, open Settings → Memory Mode; when changing the embedding model, approve the displayed plan and wait for reindexing validation.

    Tip: Choose Lite for raw conversation search without a production model, or Pro/Ultra for inference and advanced memory editing; Pro and Ultra require an active generation model.

  • Pricing / PlanWorkBuddy ve CodeBuddy

    WorkBuddy and CodeBuddy are free from November 4 to 15

    According to Tencent’s November 4 announcement, WorkBuddy and CodeBuddy are available for free from November 4 to 15. The announcement does not specify the campaign’s country, user tier, or plan eligibility.

    Usage details

    Who can use it: WorkBuddy and CodeBuddy users; campaign window 2026-11-04–2026-11-15. The official article does not explain regional or plan details.

    How to use it: During the campaign dates, sign in to WorkBuddy or CodeBuddy and start a regular work or coding session; check the billing screen in your account for pricing and eligibility.

    Tip: Try intensive tasks before November 15, but separately verify the terms on the payment screen for auto-renewal or use outside the campaign.

  • CapabilityEmoReward

    EmoReward released as open source: self-evolving rewards for emotional video RL

    EmoReward has been released as self-improving reward calibration for emotional video generation, with Plan, Act, Reflect and Evolve stages. Reflective Reward Gating uses per-sample evidence, while RewardMemory carries reliability from one training run into the next, providing a drop-in reward workflow with SAGE-GRPO and HunyuanVideo-1.5.

    Usage details

    Who can use it: Open-source training code; HunyuanVideo-1.5, VideoAlign and emotional reward provider weights are required. Linux and high GPU memory are recommended; the README recommends 32 GPUs with 80 GB-class memory per GPU.

    How to use it: After `git clone https://github.com/Tencent-Hunyuan/EmoReward.git && cd EmoReward`, run `pip install -r requirements.txt`, `bash scripts/setup_reward_subpackages.sh` and the one-command workflow `bash run_emoreward.sh`. Download the base and provider checkpoints separately.

    Tip: First generate a RewardMemory table with a short reference run; if you already have a reference run, avoid repeating the unnecessary first stage with `SKIP_STAGE1=1`.

  • APIHy-Memory

    Hy-Memory 2.0.14 adds a statuses filter for raw memory searches

    `search()` and `async_search()` now accept an optional `statuses` filter for raw L1 searches. You can request `active` and `shadow` records together; the default is `active` only, and the filter returns an error if used with non-raw searches or alongside `scene="all"`.

    Usage details

    Who can use it: `hy-memory==2.0.14` Python SDK; especially for `layers=["l1_raw"]` or Lite searches.

    How to use it: After `pip install hy-memory==2.0.14`, use `client.search("query", layers=["l1_raw"], statuses=["active", "shadow"])`. Omitting `statuses` preserves active-only behavior.

    Tip: Add `shadow` to inspect raw conversations assigned shadow status by Pro inference; keep the default to reduce noise in normal results.

  • OtherHyQEC

    HyQEC releases verifiable tools for exact quantum code distances

    The `hyqec-verify` 0.3.0 and standalone `hyqec-check` 2.0.2 packages were added to Hyra-results. The tools calculate the Pauli-weight distance of general or CSS quantum stabilizer codes using CPU/GPU search, SAT/DRAT proofs, and rerunnable certificates, and allow independent verification.

    Usage details

    Who can use it: Open-source GitHub code; requires Linux, Python 3.10 or later, a C/C++ compiler, and zlib; CUDA is optional. Licensed under Apache-2.0.

    How to use it: Create a virtual environment in the `AI4Science/hyqec` directory and run `pip install ./hyqec-verify ./hyqec-check`. Then run `hyqec-verify distance code.json --engines bnb --bnb-cert -o cert.json` and, for independent verification, `hyqec-verify verify cert.json --bnb-checker auto --threads 4`.

    Tip: To distinguish a result’s confidence level, generate a certificate and run the independent `hyqec-check` verification together; don’t present an uncertified search record as definitive proof.

  • OtherTailorRig

    TailorRig released as a rigging dataset containing 39,073 human models

    Tencent released the TailorRig dataset on Hugging Face, containing 39,073 humanoid meshes with skeletons and skinning weights. The four source-based archives can be used for auto-rigging, skinning weight estimation, and joint position regression.

    Usage details

    Who can use it: About 34 GB on Hugging Face; limited to academic research under the TailorRig License, which prohibits commercial or production use.

    How to use it: Use `huggingface_hub` to call, for example, `hf_hub_download("tencent/TailorRig", "rignet/npz_public.zip", repo_type="dataset")`, then read the `.npz` files in the ZIP. The four subsets can be downloaded separately.

    Tip: Normalize coordinates to a common scale before mixed training, and use the `body_joint_mask` field to filter out clothing and decorative bones.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 42

  • SettingQwen Code CLI

    Project hooks use a separate approval dialog

    Project hooks are now held behind an approval dialog before they run. This prevents repository-provided hooks from taking effect without user approval.

    Usage details

    Who can use it: Qwen Code v0.26.8; interactive sessions using project hooks.

    How to use it: When a project hook is triggered in an untrusted project, review the approval dialog and allow or deny it.

    Tip: Deny hooks in unfamiliar repositories; even in trusted repositories, review the hook file first.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • Pricing / PlanQwen Code

    Qwen subscription sign-in comes to CLI, VS Code and Desktop

    Qwen Code CLI, VS Code and Desktop now offer device authorization, membership checks, a model catalog, chat and built-in web search with a Qwen subscription. A shared flow has also been added for account status, plan upgrades, sign-out, and quota/limit errors.

    Usage details

    Who can use it: Qwen Code CLI Ink/OpenTUI versions, Desktop and VS Code; eligible Qwen subscription accounts and production Qwen endpoints.

    How to use it: In the CLI, select Qwen Account with Subscription from the /auth menu; in Desktop, select Qwen in the provider dialog; in VS Code, select Qwen in the provider picker. Complete browser authorization, then choose a subscription model from the model list.

    Tip: On first use, verify that the account plan and model catalog have loaded; if you run out of quota, use the /usage and /upgrade flows.

  • CapabilityQwen Code CLI

    Agent Host disconnects and unsaved joins are reported

    The CLI now prevents Agent Host disconnects from being mistakenly rolled back and reports unsaved join states. Connection issues are less likely to silently revert to an earlier state.

    Usage details

    Who can use it: Qwen Code v0.26.8; ACP/Agent Host connections.

    How to use it: Run an Agent Host-connected session with v0.26.8; watch for the status message shown when a disconnect or unsaved join occurs.

    Tip: After a disconnect, clear the unsaved state reported by the client before starting a new join.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • Plugin / IntegrationQwen Code Channels

    Bound-session action templates can be edited as needed

    Channel administrators can now edit GitHub bound-session action templates as needed. Template changes can be used for new deliveries.

    Usage details

    Who can use it: Qwen Code v0.26.8; GitHub bound-session channels.

    How to use it: In the Web Shell Channel editor, select the relevant bound-session channel, edit the action template field, and save.

    Tip: Test templates in a test repository first; clearly distinguish automatic commenting and publishing actions.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • Plugin / IntegrationQwen Code Web Shell

    Channel editor explicitly shows the event dependency for PR watch

    The Web Shell Channel editor now specifies the dependency required for PR watch events. This makes it clearer which event source the PR watch flow depends on when configuring a channel.

    Usage details

    Who can use it: Qwen Code v0.26.8; GitHub channels using PR watch.

    How to use it: Enable PR watch in the Channel editor and satisfy the displayed event dependency before configuring it.

    Tip: If PR watch does not work, first check that the channel event source and polling setting are enabled.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code daemon

    The daemon automatically discovers config Skills changes

    The daemon now automatically discovers Skills changes in the configuration. Discovery status updates faster after a Skill is added or updated.

    Usage details

    Who can use it: Qwen Code v0.26.8; installations that use the daemon.

    How to use it: Add or change the Skills source in the daemon configuration; v0.26.8 discovers the change automatically.

    Tip: Keep Skill files small and focused on a single purpose; verify loading with a test prompt after automatic discovery.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Channels

    Notifications stay unread after temporary GitHub thread errors

    GitHub notifications are no longer incorrectly marked as read when a temporary thread error occurs. The notification can still be processed again after a failed read.

    Usage details

    Who can use it: Qwen Code v0.26.8; GitHub channel notifications.

    How to use it: Use the GitHub channel with v0.26.8; if a temporary thread error occurs, reopen the notification and process it again.

    Tip: Instead of regularly clearing the notification queue, retry records that remain unread after an error.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    Gemini thought signature information is preserved during tool replay

    Gemini thought signature information is preserved across tool call replays. This reduces the risk of the model losing its earlier reasoning context in multi-turn Gemini sessions that use tools.

    Usage details

    Who can use it: Qwen Code v0.26.8; sessions using tools with the Gemini provider.

    How to use it: Upgrade Qwen Code to v0.26.8 and start a session that uses tools with the Gemini provider; preservation is applied automatically.

    Tip: In long tool chains, use context compaction when needed instead of frequently clearing the session history.

    ✓ Verified by official sourceModels: GeminiSource: Qwen Code v0.26.8 yayın notları
  • Plugin / IntegrationQwen Code Channels

    Role-aware action templates arrive for GitHub bound-session delivery

    Channel actions delivered to GitHub-bound sessions can now use role-aware action templates. Templates can select the appropriate action text based on the role in the delivery context.

    Usage details

    Who can use it: Qwen Code v0.26.8; GitHub bound-session delivery.

    How to use it: Open the GitHub bound-session channel and configure the action template settings; role information is applied automatically during delivery.

    Tip: Define separate, explicit templates for each role for consequential actions such as publishing or approval.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    Hugging Face TGI context overflow triggers automatic compaction

    Hugging Face TGI context overflow errors are now recognized and trigger the reactive compaction flow. Rather than being cut off outright, long sessions can try to compact their history when this error occurs.

    Usage details

    Who can use it: Qwen Code v0.26.8; providers using Hugging Face TGI.

    How to use it: Use the TGI provider with Qwen Code v0.26.8; reactive compaction starts automatically when the context limit is reached.

    Tip: After compaction, briefly recap important files or decisions to bring them back into focus, especially after large tool outputs.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • Plugin / IntegrationQwen Code Channels

    Replies blocked by GitHub's secondary rate limit are preserved

    When GitHub replies are blocked by the secondary rate limit, the channel flow preserves the reply without losing it. Delivery can continue more reliably after a temporary rate-limit breach.

    Usage details

    Who can use it: Qwen Code v0.26.8; GitHub channel replies.

    How to use it: Use the GitHub channel with v0.26.8; monitor the channel status to retry the reply if a secondary rate limit occurs.

    Tip: Don't set polling frequency unnecessarily high to reduce GitHub API rate limits.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code CLI

    Ink TUI tool groups use the shared loop's batch flow

    Ink TUI tool batches now run through the shared loop's batch. Tool grouping and execution order within a turn are aligned with the shared loop.

    Usage details

    Who can use it: Qwen Code v0.26.8; Ink TUI and tool calls.

    How to use it: Upgrade Qwen Code CLI to v0.26.8 and run a tool-using task in Ink TUI; the behavior is enabled automatically.

    Tip: For tasks that use multiple tools in the same turn, compare logs from the current release to monitor batch behavior.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code CLI

    Stopping Ink TUI also ends the shared tool batch

    Ink TUI's own stop action now also ends the shared tool batch. This prevents tools remaining in the shared batch from continuing after the user stops the task.

    Usage details

    Who can use it: Qwen Code v0.26.8; Ink TUI.

    How to use it: Stop the running task in Ink TUI; the current release automatically stops the shared batch.

    Tip: Check the task status to confirm that no file or remote operation remains after stopping.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code CLI

    Ink TUI model turns run in the shared agent loop

    Ink TUI model turns now run through the shared loop's turn. This means model execution in the terminal interface uses the same stopping and execution behavior as other shared-loop turns.

    Usage details

    Who can use it: Qwen Code v0.26.8; Ink TUI.

    How to use it: Upgrade the CLI to v0.26.8 and launch Ink TUI as usual; the shared loop is used automatically.

    Tip: Try the new release on a small task first to verify stopping behavior in long tool chains.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    Image input is accepted for Qwen vision models not listed in the catalog

    Image input is now accepted for vision models such as Qwen2-VL, Qwen2.5-VL, and QVQ that do not appear in the catalog. A missing modality entry in the model catalog no longer unnecessarily blocks image submission.

    Usage details

    Who can use it: Qwen Code v0.26.8; provider configurations using these model IDs.

    How to use it: Set your model provider to the relevant Qwen vision model ID and add an image to the prompt; v0.26.8 accepts the image without a catalog entry.

    Tip: Verify the model’s actual vision capabilities in the provider documentation; don’t rely on name similarity alone.

    ✓ Verified by official sourceModels: Qwen2-VL, Qwen2.5-VL, QVQSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Managed Agent

    Managed Agent can run workspace rewinds concurrently

    Managed Agent can now run multiple workspace rewind operations concurrently. This reduces unnecessary waiting between independent workspace rewind operations.

    Usage details

    Who can use it: Qwen Code v0.26.8; Managed Agent workspaces.

    How to use it: Start rewind operations for multiple workspaces in Managed Agent; v0.26.8 automatically runs eligible operations concurrently.

    Tip: Keep rewinds that touch the same file tree queued, and create a checkpoint before critical changes.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Managed Agent

    Managed Agent uses current state for approval changes

    Managed Agent now applies approval changes based on the current state rather than an outdated snapshot. This reduces the risk of overwriting the wrong state during concurrent approval updates.

    Usage details

    Who can use it: Qwen Code v0.26.8; Managed Agent approval workflows.

    How to use it: While an operation approval is pending in Managed Agent, change the permission state; the current version applies the change to the latest state.

    Tip: If multiple administrators are approving, check that the final decision matches the current state in the panel.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Managed Agent

    Managed Agent reduces shared lock wait times

    Managed Agent has narrowed the scope of shared locks across logging, output, lifecycle, workspace read/write, and session creation paths. This batch performance change aims to reduce tenant-wide wait times and bottlenecks in concurrent tasks.

    Usage details

    Who can use it: Qwen Code v0.26.8; Managed Agent multi-workspace and session workloads.

    How to use it: Upgrade Managed Agent to v0.26.8 and run concurrent workspace tasks; lock scopes are optimized automatically.

    Tip: For performance comparisons, keep the number of workspaces, tool load, and task type the same when measuring before and after.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Managed Agent

    Managed Agent explains the reason for URL health probe failures

    When a Managed Agent URL fetch fails, the health probe now gives a clearer explanation for the rejection. This makes it possible to distinguish an access issue from a rejected target URL.

    Usage details

    Who can use it: Qwen Code v0.26.8; Managed Agent URL fetch and health probe workflows.

    How to use it: Run a URL fetch or health probe in Managed Agent; the reason for the rejected request appears in the error notification.

    Tip: For security, check the URL against allowed targets and redirects separately.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • FeatureQwen Code

    Mod prompt.submit hooks run in the OpenTUI terminal

    Mod prompt.submit hooks now also run in the OpenTUI terminal, taking effect before the submitted prompt is processed. This preserves Mod-based prompt preprocessing in sessions that use OpenTUI.

    Usage details

    Who can use it: Qwen Code v0.26.8; Mods using the OpenTUI terminal.

    How to use it: Upgrade Qwen Code to v0.26.8 and run a Mod with a defined prompt.submit hook in the OpenTUI terminal; no additional enablement setting is required.

    Tip: Validate and log the prompt before changing its text in the hook to catch unexpected prompt transformations.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • AgentQwen Code

    Mods can send their own prompts in headless runs

    A Mod can now create and send its own prompt in a headless run. This enables automated Mod workflows that are not limited to a single user-written prompt.

    Usage details

    Who can use it: Qwen Code v0.26.8; headless Mod runs.

    How to use it: Define a Mod to send its own prompt in a headless run, then start the relevant headless task with v0.26.8.

    Tip: Set a repetition limit and stop condition for automatic prompt generation to prevent infinite Mod loops.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    Prompts sent by Mods run in the Ink terminal

    Prompts sent by a Mod are now actually executed in the Ink terminal. This lets additional prompts initiated by Mods join the shared session flow in the terminal user interface.

    Usage details

    Who can use it: Qwen Code v0.26.8; Mods using the Ink terminal.

    How to use it: Run a Mod that sends prompts in the Ink terminal; the prompt sent by the Mod is executed automatically.

    Tip: Keep additional prompts generated by the Mod short and focused to reduce noise in the terminal flow.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code autofix-pr

    Nightly autofix-pr helper refusals show a code and next step

    In the Nightly release, a code and next step are now shown when autofix-pr helpers are refused. Users can more quickly understand why the request was refused and how to proceed.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; /autofix-pr.

    How to use it: Run /autofix-pr in the Nightly release; if a helper is refused, check the refusal code and next step fields on the card.

    Tip: When you see the same code again, follow the recommended next step before trying again.

  • CapabilityQwen Code Channels

    Nightly GitHub inbound recovery stays active after a failed read

    The nightly release no longer disables the recovery mechanism after a failed GitHub inbound read. The channel can still recover after a temporary read error.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; GitHub inbound channel flow.

    How to use it: Use the GitHub inbound channel in the nightly release; after a read error, let the channel's automatic recovery flow retry.

    Tip: Keep error logs in the preview channel; don't confuse persistent permission errors with temporary read errors.

  • FeatureQwen Code verify-pr

    PR review threads are passed to the Nightly verify-pr sandbox run

    In the November 5 Nightly release, the sandboxed verify-pr run can now receive the PR's existing review threads. This allows validation inside the sandbox to use the context of the earlier review discussion.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; sandboxed /verify-pr.

    How to use it: Install the Nightly release and run sandboxed /verify-pr; PR review threads are passed to the sandbox run automatically.

    Tip: Before using the preview release in a production repository, try it on a read-only PR.

  • CapabilityQwen Code

    Nightly accepts image input for additional vision model families

    The Nightly release fixes image input acceptance for Llama 3.2 Vision, Llama 4, LLaVA, Pixtral, and Mistral Small 3.1/3.2. You can send image prompts to these models even if the catalog is missing a modality entry.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; relevant vision model providers.

    How to use it: Install the Nightly release, select a supported model ID, and add an image to the prompt; image input is accepted automatically.

    Tip: Check each model provider’s image format and size limits separately.

    ✓ Verified by official sourceModels: Llama 3.2 Vision, Llama 4, LLaVA, Pixtral, Mistral Small 3.1, Mistral Small 3.2Source: Qwen Code 5 Kasım nightly yayın notları
  • Plugin / IntegrationQwen Code QQBot

    QQBot proactive loop messages are delivered as active group messages

    In the Nightly release, QQBot messages sent by the proactive loop are now delivered as active group messages. Group members can see these automated messages in the regular activity feed.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; QQBot group integration.

    How to use it: Set up the QQBot group connection and enable proactive loop messages; the Nightly release sends them as active group messages.

    Tip: Keep proactive message frequency low and send messages only when there are meaningful changes for the group.

  • SettingQwen Code serve

    Conversations working-directory root can be configured for qwen serve

    qwen serve now allows you to configure the root of Conversations working directories. This lets administrators control the layout of the conversation workspaces being served.

    Usage details

    Who can use it: Qwen Code v0.26.8; installations using qwen serve.

    How to use it: Define the Conversations working-directory root option in the qwen serve configuration, then restart the server.

    Tip: Limit the root directory to only the folder the server needs to access; do not grant access to parent directories.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • APIQwen Code TypeScript SDK

    TypeScript SDK v0.1.28 released, packaging CLI 0.26.8

    Qwen Code TypeScript SDK v0.1.28 has been released. The package includes a client built from CLI v0.26.8 source code.

    Usage details

    Who can use it: TypeScript SDK users; November 5, 2026 GitHub package.

    How to use it: Upgrade the SDK package to v0.1.28 and rebuild your application’s Qwen Code client version.

    Tip: Keep CLI and SDK versions in the same release family to reduce differences in ACP and client behavior.

  • SettingQwen Code Web Shell

    Channel editor settings are grouped under Advanced settings

    Detailed Channel editor settings in Web Shell can now be collapsed under Advanced settings. This keeps the main channel editing screen simpler.

    Usage details

    Who can use it: Qwen Code v0.26.8; Web Shell Channel editor.

    How to use it: Open Channel editor in Web Shell and expand Advanced settings to access detailed channel options.

    Tip: When setting up a channel for the first time, fill in only the basic fields; adjust polling and event details later under Advanced settings.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code CLI

    Context tool details are derived from declared schemas

    The CLI now derives context tool details from the tool schemas declared at runtime. This helps the displayed or used tool information match the actual schema more closely.

    Usage details

    Who can use it: Qwen Code v0.26.8; CLI context and tool view.

    How to use it: Update the CLI and open the /context view; tool details are displayed automatically based on the defined schema.

    Tip: Keep schemas for custom MCP tools explicit and up to date; use the context view as a quick validation.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • FeatureQwen Code CLI

    Raw context usage estimates are preserved

    The CLI now preserves raw context usage estimates instead of losing them. The usage view and subsequent calculations rely on more consistent raw data.

    Usage details

    Who can use it: Qwen Code v0.26.8; CLI context usage view.

    How to use it: Open the /context command in the CLI; the current version preserves raw usage estimates automatically.

    Tip: After changing models, check the /context output again alongside the new context window.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code Mods

    Mod hooks are skipped without errors for the ui.backdrop event

    Mod hooks are now skipped when they are unsupported for the ui.backdrop event, rather than causing the entire workflow to fail. This prevents Mod incompatibility from interrupting the session for relevant UI events.

    Usage details

    Who can use it: Qwen Code v0.26.8; Mod hooks and ui.backdrop events.

    How to use it: Run the Mod with v0.26.8; if a hook is unsupported for ui.backdrop, it is skipped automatically.

    Tip: Move critical validation hooks to supported events, and monitor skipped events in the logs.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    The model retrieves the session date from history once

    The model now retrieves the date from session history once instead of generating it again on every turn. This is intended to keep date context consistent in long sessions.

    Usage details

    Who can use it: Qwen Code v0.26.8; sessions with all model providers.

    How to use it: Update Qwen Code to v0.26.8 and start a multi-turn session that uses date information; the behavior is applied automatically.

    Tip: If the day changes during a long session, start a new session to explicitly refresh the date context.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code CLI

    Session titles are preserved when Nightly rebuilds transcripts

    The Nightly release preserves titles in the session list when a transcript is rebuilt. Reprocessing history does not remove a user-provided or generated title.

    Usage details

    Who can use it: v0.26.8-nightly.20261105.7bd78683e6 preview; CLI session list.

    How to use it: In the Nightly release, change a session title and run a workflow that rebuilds the transcript; the title will be preserved.

    Tip: Save long sessions with meaningful, concise titles from the start to make them easier to distinguish.

  • APIQwen Code

    Custom model name configuration for Qwen3.x services documented

    The Qwen Code release documentation explains how to configure Qwen3.x services to be served under a custom model name. This makes it easier to match the service-side name with the client’s model selection.

    Usage details

    Who can use it: Qwen Code v0.26.8; services offering Qwen3.x under a custom model name.

    How to use it: Follow the Qwen3.x custom model name instructions in the Qwen Code v0.26.8 documentation to map the service model name to the client configuration.

    Tip: Record the custom name alongside the provider’s actual model ID; don’t confuse the two when troubleshooting.

    ✓ Verified by official sourceModels: Qwen3.xSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code verify-pr

    verify-pr fixes flaky demotion comments on closed PRs

    verify-pr now publishes flaky demotion results with the correct classification on closed PRs. Reporting inconsistencies for empty classifications and closed PRs have been fixed.

    Usage details

    Who can use it: Qwen Code v0.26.8; verify-pr and closed GitHub PRs.

    How to use it: Check a closed PR with /verify-pr; v0.26.8 reports the result with the correct flaky demotion classification.

    Tip: When reviewing automated comments, check whether the PR is open or closed alongside the report line.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • FeatureQwen Code Web Shell

    Web Shell welcome models and language are no longer session-dependent

    The model list on the Web Shell welcome screen has been aligned, and the interface language no longer depends on session state. The initial view is now more consistent in new and existing sessions.

    Usage details

    Who can use it: Qwen Code v0.26.8; Web Shell welcome screen.

    How to use it: Open Web Shell with v0.26.8 and select a model and language on the welcome screen; the selection loads independently of the session.

    Tip: After changing the language, open a new session to verify that the welcome screen and in-session language behave as expected.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • CapabilityQwen Code

    Workflow tool schemas are deferred by default

    Workflow tool schemas now load later by default. This aims to reduce unnecessary schema loading at startup and during context use.

    Usage details

    Who can use it: Qwen Code v0.26.8; workflow tools.

    How to use it: Upgrade Qwen Code to v0.26.8 and start a session that uses workflows; deferral is the default behavior.

    Tip: If you notice a delay on the first workflow call, keep in mind that tool schemas are loaded at the time of the call.

    ✓ Verified by official sourceSource: Qwen Code v0.26.8 yayın notları
  • AvailabilityQwen on Apple Intelligence

    Qwen opens up to Apple Intelligence in the iOS 27.3 developer beta in China

    According to Apple's official announcement, the iOS 27.3 developer beta in China began testing Qwen support on November 4. Users can access Qwen in certain Siri, Writing Tools, and Visual Intelligence flows; a general release is planned for iOS 27.3 in December.

    Usage details

    Who can use it: China; iOS 27.3 developer beta starting November 4, iOS 27.3 public release in December 2026; supported iPhones and English/Simplified Chinese language requirements.

    How to use it: Install the iOS 27.3 developer beta on a compatible iPhone, enable Apple Intelligence, and sign in with a Qwen account or Apple in the Qwen prompt within Apple Intelligence; then use a Qwen-powered request in Siri, Writing Tools, or Visual Intelligence.

    Tip: Before sending sensitive documents or photos in the beta environment, check your Qwen account, permissions, and data usage preferences.

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, inexpensive option at $0.10 / $0.50.

Grok 4.7

35
xAI · Grok

The default model in the Grok app. Starting November 7, Grok API will apply a default watermark to supported text models.

Tencent Hy4

35
Tencent · Hy

Correction: Hy4 is now available via API on OpenRouter (`tencent/hy4`). 50% off on TokenHub through December 31; WorkBuddy and CodeBuddy are free through November 15.

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba’s cloud flagship model; Qwen 4 has not launched yet. Qwen Code added subscription sign-in; it is being tested within Apple Intelligence in the iOS 27.3 beta in China.

DeepSeek V4.1 Flash 1017

33
DeepSeek · DeepSeek

MIT-licensed open weights, 1M context, 33 points. No new user-facing changes during this period.

Sources