YZ·RADARyz.mikail.net
Daily edition · 13 November 2026

Gemini 4 Hydrogen enters the TOP 10; Claude and Grok Bot join meetings

The TOP 10 has its first change in four days. Google’s Gemini 4 Hydrogen, released on November 12, scored 54 with the xhigh setting in Artificial Analysis Intelligence Index v5.0.1 and went straight to No. 7. As a result, GPT-6.1 Terra (51) fell to No. 8, Gemini 4 Argon (50) to No. 9, and Muse Wonder Full (50) to No. 10; Kimi K3.1 Pro (46) dropped out of the TOP 10 and moved to the top of the watchlist. Kimi remains the strongest open-weight model. Qwen3.8 Max (34) fell off the watchlist. No model’s score changed: Claude Fable 5.2 leads with 65 points, followed by Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. Hydrogen’s main advantages are price and speed. Its per-task cost on Artificial Analysis is $3,16, compared with $5,48 for Carbon; its low setting is the table’s third-fastest model, at 1,3 minutes per task. The LMArena page could not be fully read today because of a CAPTCHA; in the view that was accessible, Gemini 4 Carbon and Gemini 4 Argon lead. Qwen4-Max-Preview and Grok 4.8 still have no Artificial Analysis score.

Google’s biggest model news today was Hydrogen. The model is available through the Gemini API and AI Studio with a free tier; the paid tier costs $1,50 per million input tokens and $7,50 per million output tokens, while cached input costs $0,075. Thinking level can be set to `low`, `medium`, `high`, or the new `xhigh`; the default is `high`. The context window is 1M tokens; a single request can return up to 65.536 output tokens, with `background` execution or continuation required for longer outputs. Hydrogen also arrived in the Gemini app (with Auto model selection for free users), Google Search AI Mode, Gemini Enterprise, and BigQuery’s generative AI functions on the same day. It became the default model for the Antigravity agent in Gemini Managed Agents. Antigravity’s Orchestrator assigns planning and difficult decisions to Carbon, and routine editing and testing to Hydrogen: Orchestrator (Carbon) is available in Ultra, while Orchestrator (Argon) is available in Pro and Ultra. On the enterprise side, Agent Gateway for GKE entered preview (GKE 1.37.2-gke.1042000 and later in the Rapid channel, in europe-west1 and us-central1). The `safety_identifier` tag, which marks end users as anonymous in the Gemini API, also became generally available.

Today, two companies released meeting-joining agents at the same time. Claude Meetings entered beta for Team and Enterprise plans on November 12. For now, Claude joins only Google Meet calls as a visible participant, taking notes, answering questions, and preparing documents or whiteboards. The feature is disabled for some ZDR, HIPAA-ready, and CMEK organizations. Grok Bot 1.0, released on November 11, joins Google Meet, Zoom, and Microsoft Teams meetings and sends a summary and action items afterward (Team, Enterprise, and Ultra). With the same release, Team Bots can work in Microsoft Teams channels, connector logins can be shared with the team using the admin-enabled Allow Shared Logins setting, and what team members teach the Bot awaits approval in Review learnings. Anthropic’s 53 entries go beyond that. The updated Usage Policy and Supported Regions Policy took effect on November 12. Claude for Manufacturing, an extension for production teams that includes 17 skills and connectors for Autodesk Fusion, Infor, Microsoft 365, and Oracle NetSuite, was released. Claude Code 2.1.323 introduced experimental Code Mode, which uses JavaScript for tool calls. The default context for unrecognized model IDs increased from 200K to 1M, and the legacy MCP v1 client was removed. Claude Desktop v5.11492.0 adds ADMX/Profile Manifest template export for admins, client certificates tied to the allowed-host list, and up to 5 concurrent sub-sessions in Auto mode.

There were no model or pricing changes at OpenAI. Codex CLI 0.168.0 lets users choose the default model provider for new chats, open a second chat alongside the first with experimental `tui_workspace`, attach audio recordings, and generate animated GIFs with a configuration flag. The Python, Node, Go, and Ruby SDKs now return reasoning input transformations in Responses responses as typed values. Alibaba launched AI Gateway Intelligent Routing: set `auto/balanced`, `auto/cheap`, `auto/fast`, or `auto/quality` for the model to select the appropriate model for each request (currently on dedicated instances). Qwen3Guard can be connected to the Higress gateway as a Wasm plugin. AgentScope Service was released as an open-source control plane, and Cloud Monitor 2.0 brought observability for agents. Qwen Code 1.0.0-preview.1 and Web Shell include more than 20 changes. The most urgent of Tencent’s 48 entries is this: Hy-Image-3.0 (`hy-image-v3`) was removed from TokenHub on November 12 at 23:59 Beijing time, with `hy-image-v3.5-preview` recommended as its replacement. BrowserSkill gained plan and recipe flows, Hy-Memory 2.0.16 gained PostgreSQL support, and DSpine gained Apple Silicon (MLX) support; LoopForge and teamai-cli were also updated. At Moonshot, Kimi Code Desktop v1.3.4 and 2.6.0 canary were released. At Meta, the Threads Trending Topics API became available to all developers (US and Japan trends), and `call_permission_mode` was added to the WhatsApp Cloud API. There are no new entries on Z.ai’s official channels; SambaNova added GLM-5.3 to its SambaCloud platform. DeepSeek and Xiaomi had no new changes. A total of 200 new developments were added to the archive today.

Today’s TOP 10

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

  1. 01

    Claude Fable 5.2

    Anthropic · Claude=
    65pts

    Released on November 9 and leads with 65 points. Available on all paid Claude plans; API model `claude-fable-5-2`, $8 / $40, with thinking always adaptive. The Claude Meetings beta (Google Meet first) opened to Team and Enterprise plans on November 12.Released: 9 November 2026

  2. 02

    Gemini 4 Carbon

    Google · Gemini=
    60pts

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

  3. 03

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

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

  4. 04

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

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

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

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

  6. 06

    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

  7. 07

    Gemini 4 Hydrogen

    Google · GeminiNEW
    54pts

    Released on November 12 and entered directly at No. 7 with 54 points using the xhigh setting. Its per-task cost on Artificial Analysis is $3,16 ($5,48 for Carbon). API model `gemini-4-hydrogen` is also available on the free tier; paid pricing is $1,50 / $7,50. Available in the Gemini app, Search AI Mode, Gemini Enterprise, and Antigravity (Pro/Ultra).Released: 12 November 2026

  8. 08

    GPT-6.1 Terra

    OpenAI · ChatGPT▼ 1
    51pts

    Released on ChatGPT Work, Codex, and the API on October 26; focused on vision and computer use. Dropped one place after Gemini 4 Hydrogen entered. The admin preview for Dots agents runs through November 18.Released: 26 October 2026

  9. 09

    Gemini 4 Argon

    Google · Gemini▼ 1
    50pts

    Available on Google AI Pro/Ultra and the paid API; promotional pricing of $2 / $10 through November 30. Dropped one place after Hydrogen entered. Orchestrator (Argon) is available to both Pro and Ultra subscribers in Antigravity.Released: 30 September 2026

  10. 10

    Muse Wonder Full

    Meta · Meta AI▼ 1
    50pts

    Meta’s most powerful Muse model; no public API. Fell one place to No. 10. The Muse app is also available on Apple Watch and Wear OS.Released: 12 October 2026

Key signals

  • The TOP 10 changed: Gemini 4 Hydrogen entered at No. 7 with 54 points; Terra, Argon, and Muse Wonder Full each fell one place, and Kimi K3.1 Pro (46) dropped off the list. Claude Fable 5.2 (65) is at the top.
  • Google: Hydrogen is available with a free tier in the API ($1,50 / $7,50), as well as in the Gemini app, Search AI Mode, and Antigravity. Antigravity Orchestrator splits work between Carbon and Hydrogen.
  • Meeting agents: Claude Meetings beta (Team/Enterprise, Google Meet) and Grok Bot 1.0 (Meet, Zoom, Teams) join meetings and take notes.
  • Anthropic: Experimental Code Mode arrived in Claude Code 2.1.323, and the MCP v1 client was removed. Usage Policy and Supported Regions Policy took effect on November 12.
  • Urgent migration: Tencent Hy-Image-3.0 was removed from TokenHub; migrate code using `hy-image-v3` to `hy-image-v3.5-preview`.
  • 200 new developments: Anthropic 53, Tencent 48, Alibaba 30, xAI 19, OpenAI 17, Google 16, Moonshot 12, Meta 4, Z.ai 1. No new changes at DeepSeek or Xiaomi.

Worth trying today

01

Cut costs by assigning routine tasks to Gemini 4 Hydrogen

Select Gemini 4 Hydrogen in the model selector in Google AI Studio, or set the model to `gemini-4-hydrogen` in the API; the free tier is sufficient for prototyping. For quick tasks such as classification, summarization, and short code fixes, set `generation_config.thinking_level` to `low`; use `xhigh` for difficult analysis. Since the default is `high`, you may spend more thinking tokens than necessary if you do not specify a level. If you send the same long system prompt repeatedly, keep the shared prefix fixed; cached input costs $0,075 per million tokens. Use `background=true` if you need an output longer than 65.536 tokens. If you use Antigravity, select Orchestrator from the model menu so Carbon handles planning and Hydrogen handles routine editing.

Google · Gemini
02

Leave meeting notes to Grok Bot or Claude Meetings

In Grok Bot (Team, Enterprise, or Ultra), paste a Google Meet, Zoom, or Microsoft Teams link into the Bot chat and say “join and take notes.” Approve the confirmation card, and have the meeting host admit the Bot; a summary and action items will arrive when the meeting ends. A Bot that follows your calendar can join scheduled meetings on its own; manage these under Upcoming in the Bot details. In Enterprise, an admin must first enable Meetings. If you use Claude Team or Enterprise, add Claude to a Google Meet meeting from the Claude.ai calendar or enable automatic joining in Meetings settings. In both cases, tell participants in advance that an AI is taking notes.

xAI · Grok
03

Check three settings after upgrading to Claude Code 2.1.323

After updating, verify that you are on 2.1.323 with `claude --version`. First, if you use a model Claude Code does not recognize through Bedrock, Vertex, or another provider, the default context is now 1M; if your model does not support that much, set `CLAUDE_CODE_DISABLE_1M_CONTEXT=1`. Second, `MCP_SDK_GENERATION=v1` now runs the v2 compatibility mode rather than the legacy client; test your old MCP servers once, especially those accessed through OAuth. Third, experimental Code Mode is optional; an admin can disable it in settings if your organization does not allow it. `/security-review` now asks for confirmation before running git commands in risky directories, so run it on a normal clone.

Anthropic · Claude
04

Move Hy-Image-3.0 calls to Hy-Image-3.5-preview immediately

The `hy-image-v3` model parameter in Tencent Cloud TokenHub stopped working on November 12 at 23:59 (Beijing time). Search your codebase for the old value (`grep -rn "hy-image-v3\b" .`) and replace it with `hy-image-v3.5-preview`. Compare the output using a few sample prompts, check billing, and then deploy, following the Hy Image Generation API guide. The model is still in preview; keep a backup provider ready if image generation is critical to your workflow.

Tencent · Hy

This edition’s updates by AI

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

Anthropic · Claude

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

New in this edition · 53

  • AgentClaude Code agents

    Agent teams and workflows preserve leader, approval, and usage-limit states

    Agents in split panes now verify that requests come from the team leader; if the leader has ended, the request is rejected, and plan approval is requested again after `/clear` or `/resume`. Desktop Code tab workflows pause at usage limits and resume when the limit resets; agents stopped by the safeguard are prompted to retry with a narrower prompt.

    Usage details

    Who can use it: Claude Code agent teams, workflows, and Desktop Code tab.

    How to use it: Keep the leader session open; approve the plan again after `/clear` or `/resume`. Wait on the limit card or switch to an eligible model from the options menu.

    Tip: For long workflows, check when the limit resets and save critical intermediate outputs outside the workflow as well.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Background and Remote Control sessions fixed for identity drift and TLS disconnections

    Sessions resumed in the background can no longer switch to a different login instead of a named profile, API key, or workload identity; if a profile is missing, the session stops and explains why. Login, token refresh, and logout no longer wait indefinitely on unresponsive TLS handshakes; when a gateway login is refreshed, the next message continues.

    Usage details

    Who can use it: Claude Code terminal, background sessions, Desktop, and Remote Control.

    How to use it: Upgrade to 2.1.323; reopen active sessions after changing profiles. If a gateway session expires, run `/login` in another terminal.

    Tip: To prevent identity drift in automations, pin a named profile or explicit API identity and log warnings from background sessions.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Additional approval boundaries added for artifact updates and app-originated prompts

    Artifact updates that may be visible to people outside the organization, or whose visibility cannot be verified, now require confirmation even when an allow rule exists. Prompts that the app marks as “not typed” are no longer passed to agents as user requests; `/skill` or `@file` references among selected files are no longer applied accidentally.

    Usage details

    Who can use it: Claude.ai, Cowork, Desktop, and Claude Code artifact/workflow flows.

    How to use it: Check artifact visibility and approve external-sharing warnings; enter file/skill calls on a separate prompt line.

    Tip: Use explicit human approval rather than an allow rule in automations that expose artifacts externally.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AgentClaude Desktop Code

    Sessions started in Auto mode can open sub-sessions for a task

    In a local Code session, a session started by Claude can open new sessions for a task if the Auto mode safety check allows it. The limits are 10 per message you write and 5 at a time.

    Usage details

    Who can use it: Claude Desktop v5.11492.0 local Code sessions; excludes scheduled runs.

    How to use it: Enable Auto mode in a Code session and give it a task; if the safety check allows it, Claude will start sub-sessions automatically.

    Tip: Parallel sub-sessions can increase file conflicts; divide tasks among independent folders or worktrees.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • CapabilityClaude Code

    Bypasses in Bash, Write/Edit, sandbox, and permission hook checks have been closed

    The possibility of running without requesting permission through `sudo`, `env`, wildcards, zsh options, `cp`/`mv`, `rm`, `lsof`, and some `sandbox.excludedCommands` paths has been fixed. Write/Edit tools can no longer bypass Read deny rules, the sandbox can no longer be disabled during a session, the plan file can no longer be written outside the project, and incorrect decisions based on permissions changed by a hook have been prevented.

    Usage details

    Who can use it: Claude Code across all interactive, `-p`, SDK, Auto, and bypass flows.

    How to use it: Update to 2.1.323; retest your permission rules using full commands, wrappers, and actual target paths.

    Tip: Use narrowly scoped allow/deny rules and `--permission-mode auto` in production instead of bypassing permissions; do not silently skip security prompts through automation.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Code

    Claude Code adds experimental Code Mode

    Code Mode has been added as an experimental, opt-in mode in which Claude writes JavaScript for tool calls instead of calling its tools directly. Organization admins can disable the mode at the organization level.

    Usage details

    Who can use it: Claude Code 2.1.323 and later; experimental and opt-in.

    How to use it: Update to Claude Code 2.1.323 and apply the activation setting in the official Code Mode instructions; organization admins can disable the mode in managed settings.

    Tip: Try it first in a low-risk repository with limited tool permissions; verify that JavaScript-based calls are subject to your usual permission rules.

  • FeatureClaude Desktop

    Claude Desktop exports ADMX and Profile Manifest configuration templates

    The Developer > Export Configuration Template menu generates ADMX and Profile Manifest policy templates for standard Claude Desktop, third-party inference, or both.

    Usage details

    Who can use it: Claude Desktop v5.11492.0; administrators.

    How to use it: In Claude Desktop, open Developer > Export Configuration Template, select the target configuration type, and export the template.

    Tip: After adapting the template to your organization’s policies, deploy it to a small test group in stages; older clients may reject unknown keys.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • Plugin / IntegrationClaude for Manufacturing

    Claude for Manufacturing plugin launches with 17 skills and four connectors

    The new plugin for manufacturing teams supports workflows such as preparing quotes, quality reports, troubleshooting, and supplier requests with 17 skills. It works with Autodesk Fusion, Infor, Microsoft 365, and Oracle NetSuite connectors, and completes actions with user review/approval.

    Usage details

    Who can use it: Available as a Claude Marketplace plugin on Team and Enterprise plans.

    How to use it: In Claude, install the plugin from Customize > Plugins and type “Set us up” in the chat; Claude configures the systems and connectors and reports which skills can run.

    Tip: For first-time use, connect access to manufacturing records through least-privilege accounts, and always require human approval for quotes or purchasing changes before they are published.

    ✓ Verified by official sourceSource: Introducing Claude for Manufacturing
  • AvailabilityClaude Meetings

    Claude Meetings rolls out in beta to Team and Enterprise plans

    Claude Meetings has begun rolling out in beta within Claude and Claude Desktop for Team and Enterprise plans. It can join Google Meet calls as a visible participant, take notes, answer questions, and create documents, whiteboards, or artifacts.

    Usage details

    Who can use it: Team and Enterprise plans; initially, Google Meet. Unavailable to some ZDR, HIPAA-ready, and CMEK organizations.

    How to use it: Add Claude to a meeting from the Claude.ai calendar or turn on automatic joining in Meetings settings; an Enterprise admin must first enable the feature for the organization or relevant group.

    Tip: Let everyone in the meeting know in advance that it will be transcribed; note that responses using connected tools may count toward plan usage.

  • APIClaude Tag

    Slack sessions and more granular admin controls arrive for Claude Tag

    In the Enterprise beta, Claude Tag sessions started from Slack can now be listed and their transcripts read by adding them to the Compliance API. Admins can use a confirmation dialog before turning the feature off, Retry if channel settings fail to load, and the Scope filter under Activity > Network events.

    Usage details

    Who can use it: Claude Enterprise; Claude Tag and Compliance API beta features.

    How to use it: Enable Claude Tag in Enterprise administration; query Slack sessions in the Compliance API. Select Scope under Admin > Claude Tag > Activity > Network events.

    Tip: Reduce review costs by narrowing compliance exports to the relevant scope rather than the entire organization.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code on the web

    Reconnection fixed for Cloud, self-hosted, and Remote Control sessions

    Loss of messages, large images, subagent responses, and streamed text in restarted cloud sessions has been fixed; self-hosted sessions getting stuck in Starting/Sending have also been fixed. `Try again` on the connection card now attempts to reconnect instead of sending the words as a user message; web sessions can reconnect automatically.

    Usage details

    Who can use it: Claude Code on the web, self-hosted runner, and Remote Control.

    How to use it: When a session disconnects, click `Try again`; reopen old sessions after updating and, if needed, monitor the status on the connection card.

    Tip: For long tasks, write intermediate outputs to the repository; even if the connection is restored, critical results won’t depend on chat history.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Compaction, resume, transcript, and large-file session recovery fixes released

    Fixed lost messages that Claude hadn't responded to, or new messages being attached to an old interruption, during repeated `/compact`, `/context`, `/resume`, and worktree switches. Resolved memory spikes during large file changes, separate saving of subagent transcripts/tool results, and CLI, SDK, and VS Code sessions getting stuck after low-memory events.

    Usage details

    Who can use it: Claude Code CLI, SDK, Desktop, VS Code, and cloud sessions.

    How to use it: Upgrade to 2.1.323; close and reopen active sessions during the update, then check history with `/resume` after recovery.

    Tip: Make large edits in separate commits; this makes it easier to restore file state even if there's a transcript recovery issue.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Cowork

    GitHub and package registry write operations blocked in Cowork restricted network tasks

    Tasks with restricted network access can only clone and fetch from GitHub; they cannot push, use Git LFS, or use SSH remotes. Downloads and npm audit remain available for `registry.npmjs.org`, `registry.yarnpkg.com`, and `crates.io`, while publishing, login, and other requests are rejected.

    Usage details

    Who can use it: Cowork tasks with restricted network access; those with Domain allowlist set to All domains and third-party deployments are unaffected.

    How to use it: Check the task’s network policy; if publishing is required, use an explicitly permitted deployment environment or the All domains policy.

    Tip: Have a separate CI/CD step publish to production after review, rather than publishing from the Cowork task itself.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • OtherClaude Usage Policy

    Updated Usage Policy takes effect and clarifies new usage limits

    The policy in effect as of November 12 adds a separate prohibition on deceptive campaigns and artificial engagement, and clarifies rules for undermining elections, weapons software and autonomous weapons actions, non-consensual surveillance, and law enforcement decisions. Human oversight and notifying affected people that AI is being used remain requirements for high-risk use; stoppability and safe-state requirements have been added for hardware that performs physical actions.

    Usage details

    Who can use it: All users of Claude.ai, Claude Code, API, developer platforms, cloud providers, and authorized resellers.

    How to use it: Before production, reassess your use case against the Universal Usage Standards, High-risk Use Case Requirements, and Additional Use Case Guidelines sections at https://www.anthropic.com/legal/aup.

    Tip: For workflows using agents or physical hardware, document that a human can intervene and that the system enters a safe state if the connection is lost.

  • Plugin / IntegrationClaude Desktop Code

    Organization connectors set to Needs approval now require approval in Code sessions

    Local, SSH, and WSL Code tab sessions now require approval for connector tools set to Needs approval in the claude.ai organization policy, regardless of permission mode. In workflows other than Auto and plan mode, approval may offer the option to grant permission for the session.

    Usage details

    Who can use it: Claude Desktop Code; excludes scheduled runs and third-party deployments.

    How to use it: Accept the approval prompt when a connector tool is called; if you need ongoing permission, select `Allow for this session`.

    Tip: Scheduled tasks cannot receive human approval, so set connectors to a policy appropriate to the task's risk instead of Needs approval.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • APIClaude Code MCP

    MCP SDK v1 client removed; v1 variable now enables v2 compatibility mode

    `MCP_SDK_GENERATION=v1` no longer runs the old v1 client; it runs the v2 MCP client in compatibility mode, with an earlier handshake and relaxed OAuth input checks.

    Usage details

    Who can use it: Claude Code MCP clients and older MCP servers.

    How to use it: Keep `MCP_SDK_GENERATION=v1` temporarily for older servers; test the v2 handshake and OAuth flow with newer servers.

    Tip: Don't treat compatibility mode as a permanent solution; move your MCP server to the v2 handshake and current OAuth expectations.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code MCP

    User approval now required to complete MCP server sign-in

    The tool that completes MCP server sign-in now requires approval even in bypass and Auto mode. In `dontAsk` flows, where the user cannot be prompted, `/mcp` must be opened to sign in.

    Usage details

    Who can use it: Claude Code MCP server sign-ins.

    How to use it: Accept the approval prompt shown when connecting an MCP server; in `dontAsk` mode, open `/mcp` and complete the sign-in flow there.

    Tip: If you use `dontAsk` in automations, complete MCP authentication in an interactive session first.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code plugins

    Security and visibility safeguards expanded for plugin and marketplace installs

    Marketplace/plugin installation fails if a git submodule is not at the repository's recorded commit; if a plugin command writes to stderr, a warning is shown instead of success. claude.ai conversations and Cowork tasks check plugin files for potential secret keys and ask before sending them; errors in plugins enabled through managed settings are now visible in SDK/Desktop.

    Usage details

    Who can use it: Claude Code plugin, marketplace, claude.ai, and Cowork flows.

    How to use it: Install the plugin with `claude plugin install` or `/plugin`; review warnings. Read managed plugin errors from the `init.plugin_errors` field in the SDK.

    Tip: Pin marketplace commits and treat stderr output as an error in CI before distributing plugins.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Boundaries for proxy, managed env, and autocompact settings clarified

    Organization settings and policy requests no longer use a proxy only through `ALL_PROXY`; `HTTPS_PROXY` is required. Managed `policyHelper` does not take environment variables from user/`--settings` files; `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE` no longer applies to auto-memory and some `/btw` flows in hosted cloud.

    Usage details

    Who can use it: Claude Code enterprise, hosted cloud, and managed settings.

    How to use it: Set `HTTPS_PROXY` for the proxy; move `policyHelper` requirements to managed `env` or the launching shell.

    Tip: Test organization-wide proxy and helper changes with a single test account first; don’t rely on user settings.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Sandbox, symlink, worktree, and sensitive permission rules tightened

    The Bash sandbox now rejects symlinks leading out of writable directories, git worktree operations in other sessions' worktrees on macOS/Linux, and some unsafe `glab` operations. “Don't ask again” rules that appear to contain secrets or would write to a tracked settings.local.json are now kept only for the session; startup stops when sandbox settings are invalid.

    Usage details

    Who can use it: Claude Code on macOS, Linux, and WSL; all sessions with security policies.

    How to use it: Add the actual target path to the allowlist; perform worktree operations in the worktree where the session started. Fix invalid `--settings` JSON and restart.

    Tip: Write Allow rules using the actual command and target path rather than a wildcard; test `.git`, `.claude`, and symlink targets separately.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    security-review and inline git operations now require approval first

    `/security-review` now requires approval before running inline git commands in the skill in risky directories such as bare repository worktrees; this restriction applies to flows outside auto mode and bypass.

    Usage details

    Who can use it: Claude Code 2.1.323; terminal, SDK, and related skill flows.

    How to use it: Run `/security-review` or the relevant skill and accept the approval prompt; use a normal clone to run it in a safe directory.

    Tip: Instead of removing the approval step when automating security reviews, run the repository in a separate, disposable worktree.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude ve Anthropic platformları

    Supported Regions Policy clarifies coverage for use in unsupported regions

    Effective November 12, the policy explicitly excludes use by people physically located in unsupported regions and by organizations established or headquartered in unsupported regions. The same rule also applies to organizations directly or indirectly majority-owned or controlled by people or organizations in those regions.

    Usage details

    Who can use it: Anthropic apps, API, developer platform, Claude models on cloud platforms, and authorized resellers.

    How to use it: Compare the organization's physical location, country of establishment, and ownership/control structure with the list at https://www.anthropic.com/legal/supported-regions.

    Tip: For access through cloud or resale, include the legal entity and ownership status in compliance checks, not just the user's IP location.

  • SettingClaude Desktop third-party deployments

    New managed setting keys added for third-party deployments

    Configurations for `autoModeClassifierModel`, `deploymentLogo`, `enforceConsoleSignInOrgList`, beta `remoteFileDownloads`, and Microsoft Purview `sensitivityLabels` have been added. Mermaid code block and Markdown previews, along with Web search/Web fetch controls in the Connectors menu, are also available.

    Usage details

    Who can use it: Claude Desktop third-party and Enterprise-managed deployments; some keys are available only in device profiles or beta.

    How to use it: Get the latest template using Developer > Export Configuration Template; add the relevant keys to the managed configuration and update clients. For Connectors, open + > Connectors or Customize > Connectors.

    Tip: Upgrade all devices to v5.11492.0 before adding new keys to the MDM profile; older Linux/Windows versions may reject unknown keys.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • FeatureClaude Code for VS Code

    VS Code trust prompt and live Bash/PowerShell output added

    Before the first session in an untrusted folder, VS Code now displays a Claude Code trust prompt; declining disables folder settings and MCP servers. Live output from a running Bash or PowerShell command also streams into the chat below the command card.

    Usage details

    Who can use it: Claude Code VS Code extension, 2.1.323.

    How to use it: Open a folder in VS Code and start your first Claude Code session; trust the folder. While a command is running, watch its output below the command card in the chat.

    Tip: Decline to trust repositories whose source you have not reviewed to prevent project settings and MCP servers from being used automatically.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Desktop Code

    Host MCP servers run under Claude Code with built-in SSH

    On Mac and Linux hosts, the built-in SSH library now runs MCP servers in `claude_desktop_config.json` under Claude Code on the host when `sshTransport: builtin` is used. Servers follow host MCP rules and stop when the session ends.

    Usage details

    Who can use it: Claude Desktop Code; built-in SSH, Mac/Linux host.

    How to use it: Set `sshTransport` to `builtin` in the managed configuration and update the host’s MCP servers in `claude_desktop_config.json`.

    Tip: Verify which user and PATH the MCP commands run with on the host; host permissions differ from those on the local machine.

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

    Built-in browser restricts client certificates to an allowed-host list

    The built-in browser and preview panels now send client certificates only to sites listed in device management's `builtinBrowserClientCertificateHosts` list. A site that requests a certificate but is not on the list will not load; `*` can be used for a proxy that requests a certificate on every site.

    Usage details

    Who can use it: Claude Desktop v5.11492.0; device-managed deployments.

    How to use it: Add `builtinBrowserClientCertificateHosts` to the managed configuration and enter the required hostnames; use `*` if a certificate is required for every site.

    Tip: Using a narrow host list instead of `*` prevents client certificates from being sent to unnecessary sites.

  • FeatureClaude Code Artifact tool

    .docx support added for Artifact tool asset uploads

    The Artifact tool can now upload `.docx` files as assets.

    Usage details

    Who can use it: Claude Code 2.1.323 and later.

    How to use it: Open the file upload option in the Artifact workflow and select a `.docx` file, or call the asset upload through a plugin/skill.

    Tip: For large documents, extracting the relevant sections before uploading reduces context usage.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Cowork

    File downloads added for Cowork tasks running on your computer

    A Download option was added to the menu for files from tasks running on your computer; a copy of the file can be saved to a location you choose.

    Usage details

    Who can use it: Claude Desktop v5.11492.0; Cowork tasks running on your computer.

    How to use it: Open the file menu in the task output, select `Download`, and choose a destination folder.

    Tip: Preserve traceability by adding the source task and version information to the downloaded file’s name or alongside it.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • CapabilityClaude Code

    CLI session deletion, logout, and usage error messages are clearer

    `claude rm` and the agents view now say that they remove a conversation from the list rather than deleting it; the conversation remains available with `claude --resume`. `/logout` in a background session now logs out all sessions on the computer and deletes MCP logins and plugin secrets; cloud/Remote Control limit errors and `/usage-credits` no longer provide a web address.

    Usage details

    Who can use it: Claude Code CLI, background sessions, and cloud/Remote Control.

    How to use it: Use `claude --resume` to continue a conversation; confirm the impact before running `/logout` in the background to log out of all sessions.

    Tip: Do not run `/logout` on an automation account; it also clears credentials for other sessions.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code on the web

    Fixes for cloud repositories, GitHub Enterprise, Actions logs, and scheduled task flows

    GitHub Enterprise Server PR updates now connect to their own server; very large GitHub Actions logs can be downloaded. New cloud sessions opened with a repository from iOS now get a separate branch, long clone/fetch operations have more resilient timeouts, and a save check prevents the GitHub trigger from being moved incorrectly when a repository is removed.

    Usage details

    Who can use it: Claude Code on the web; GitHub Enterprise Server, iOS, and scheduled tasks.

    How to use it: Check the repository and branch when opening a cloud session; reconnect GitHub Enterprise using your own host.

    Tip: Before removing a repository from a scheduled task, disable the trigger or don’t leave the page without saving the selection of a new repository.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureCode Review

    CSV cost export added to Code Review admin settings

    A button has been added to Code Review admin settings to export review costs per repository as a CSV by day, week, or month. Reports can be generated for a single repository, a GitHub organization, or the entire organization.

    Usage details

    Who can use it: Code Review admin settings.

    How to use it: Open Code Review admin settings, select the period and scope, and click the `Download CSV` button.

    Tip: Start by exporting repository-level data; expand to the organization after identifying anomalies.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Pricing / PlanClaude Code

    Reset time and options are clearer for Fable and spending limits

    Fable limit messages in `-p`, SDK, IDE, and cloud sessions show the reset time and options for continuing. `/usage` now preserves available limit bars in sessions using a setup token or `CLAUDE_CODE_OAUTH_TOKEN` and explains which ones are missing; the usage credit prompt also appears in the app in Remote Control.

    Usage details

    Who can use it: Claude Code terminal, SDK, IDE, cloud, and Remote Control sessions.

    How to use it: Run `/usage`; check the reset time on the limit card and choose to wait or switch from the usage credit menu.

    Tip: When the Fable limit is reached, keep in mind that even small side requests such as naming can consume quota.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude apps gateway

    Model, MCP, and logging behavior improved in Gateway and OpenTelemetry output

    Gateway `/v1/models` returns `max_input_tokens` for models with a definite context window; OpenTelemetry reports MCP names according to the current connector directory. Fixes were made to JSON log escaping, proxy timeout, the default OTLP protocol, and preserving other fields when resource attribute input is malformed.

    Usage details

    Who can use it: Enterprise deployments using Claude apps gateway and OpenTelemetry.

    How to use it: Upgrade the Gateway to 2.1.323; test your `/v1/models` response and OTLP collector with the new `max_input_tokens` and corrected JSON/HTTP protobuf behavior.

    Tip: On the collector side, combine `max_input_tokens` with model capacity and check context before making a request.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Desktop

    Import, managed settings, and authentication resilience improved

    Claude (third-party) imports now preserve scheduled tasks, projects without sessions, and accepted instructions; session, project, and task remnants are cleared when copying files or folders. Slow sign-in refreshes now offer a retry without losing the message; local managed `cleanupPeriodDays` is applied when needed.

    Usage details

    Who can use it: Claude Desktop v5.11492.0; third-party import and managed configurations.

    How to use it: Run the Import as my own flow again; upgrade devices to v5.11492.0 and put the managed cleanup setting in the device or server configuration.

    Tip: After importing, review tasks while they are closed and verify the correct identity and folder permissions.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • Pricing / PlanClaude Code

    An options menu is now available when usage credits are disabled or depleted

    If model usage credits are disabled or depleted, Claude Code now displays a menu with options to stop, switch to usage credits, or upgrade.

    Usage details

    Who can use it: Claude Code 2.1.323; sessions that hit usage credit or plan limits.

    How to use it: When you reach the model limit, select `Stop`, `Switch to usage credits`, or `Upgrade` from the menu that appears.

    Tip: Before starting critical, lengthy tasks, use `/usage` to check when your Fable and session limits reset.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code MCP

    MCP approvals now work more consistently with exact server names and tool schemas

    Project `.mcp.json` approvals now match the server's exact name; differently written names require approval again. When `/mcp` excludes schemas the API would reject, it shows how many tools were excluded and why; ToolSearch also reports tools with short names that are already available.

    Usage details

    Who can use it: Claude Code MCP and ToolSearch.

    How to use it: Match the server name in `.mcp.json` exactly to the name in the MCP registration; check the excluded-tool counts in `/mcp` and ToolSearch output.

    Tip: Using a single canonical name instead of generating server names automatically reduces repeated approval prompts.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code plugins

    Mod hook and plugin tool scope are now more clearly defined

    Wildcard/namespace/negated hooks on `engine.create` now only observe; named hooks are required to add or retain nouns. Tools added or removed with `tool.list` now apply to the model and all agents; invalid engine.create formats are rejected by `plugin validate`.

    Usage details

    Who can use it: Claude Code mod and plugin developers.

    How to use it: Separate mod hooks into named `engine.create` hooks and validate with `claude plugin validate --strict`.

    Tip: Run regression tests on plugins that rely on the old wildcard hook behavior before upgrading.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Cross-session message limit raised to 100,000, and unknown models now assume 1M

    The number of messages sessions on the same machine can send to one another without a user prompt has increased from about 20 to 100,000; the same limit now applies to the Desktop Code tab. The default context for model IDs Claude Code doesn't recognize has increased from 200K to 1M.

    Usage details

    Who can use it: Claude Code 2.1.323; includes sessions with third-party providers for unrecognized model IDs.

    How to use it: Use `/autocompact` for earlier compaction; set `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` to revert to the 200K assumption for unrecognized models.

    Tip: Set this environment variable in advance if your provider's actual context window is smaller than 1M; otherwise, `/clear` may be required in a full session.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code plugins and MCP

    Plugin hot reload, MCP reconnects, and connection errors are more informative

    Fixes address reauthentication prompts after MCP sign-in expires, tools not appearing after reconnect, old calls being lost when a server connects late, and hook responses over 16 MiB being silently skipped. Errors involving plugin hot reload, marketplace renames, disabled/default plugins, and long Windows paths are now handled more visibly or reliably.

    Usage details

    Who can use it: Claude Code plugin, MCP, Agent SDK, and `claude -p`.

    How to use it: Select Reconnect from `/mcp`; after installing a plugin, check `/plugin` and the reasons shown in the output of `claude plugin list`.

    Tip: Keep MCP hook output under 16 MiB and choose stable, unique server and marketplace names.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code plugins

    Plugin version compatibility is now governed by the engines field in plugin.json

    The engines["claude-code"] field in plugin.json determines the range of Claude Code versions in which the plugin can be installed, updated, or loaded. Installation and updates select the newest version compatible with the range.

    Usage details

    Who can use it: Claude Code 2.1.323 and later; plugin developers and administrators.

    How to use it: Add the supported Claude Code version range to your plugin's plugin.json under engines["claude-code"], then update with `claude plugin install` or `/plugin`.

    Tip: To avoid breaking changes, narrow the lower and upper version bounds to the versions you have tested.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code plugins

    plugin validate adds TypeScript hook type checking and mod noun reporting

    `claude plugin validate` type-checks the plugin's TypeScript hooks module; findings produce warnings, while `--strict` causes failure. It also reports `$` nouns added, retained, or changed by the mod's `engine.create` hooks.

    Usage details

    Who can use it: Claude Code plugin developers; 2.1.323 and later.

    How to use it: Run `claude plugin validate`; use `claude plugin validate --strict` in CI to turn warnings into errors.

    Tip: Add the `--strict` check to CI before release to catch incompatibilities in the hook API early.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Improvements to plugin-heavy sessions, installation, cleanup, and WebFetch performance

    Startup delays are shorter and missing-plugin reasons are now visible in claude.ai/Desktop sessions with more than 2,000 plugins enabled; the install script no longer downloads the same version twice. WebFetch is faster and uses less memory on deep pages, cleanup treats unreadable folders as errors, and `CLAUDE_CODE_AUTOMOUNT_ROOTS` can be used when needed for Linux autofs.

    Usage details

    Who can use it: Claude Code Desktop, cloud, Linux, and plugin-heavy installations.

    How to use it: Update the install script; set `CLAUDE_CODE_AUTOMOUNT_ROOTS` if the autofs root is multiple levels deep.

    Tip: Using project-based activation instead of enabling many plugins reduces startup time and the error surface.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code Remote Control

    Remote Control reconnects, permission display, and device identity have been fixed

    Fixes address connection drops after Wi-Fi/VPN changes, pending terminal permissions not appearing on mobile, slots not being freed when the disk stops, and sessions incorrectly appearing to be running. A host restarted from the same folder after more than four hours now preserves the same computer identity for up to 30 days.

    Usage details

    Who can use it: Claude Code Remote Control and mobile/web clients.

    How to use it: Restart the Remote Control host with the latest version; open the permission prompt on mobile and complete the step awaiting a response in the terminal.

    Tip: Restarting the host from the same folder preserves device pairing; avoid switching folders to create a new identity.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code on the web

    No repository option added for new sessions in self-hosted environments

    An option has been added to start without linking a repository when opening a new Claude Code web session in a self-hosted environment.

    Usage details

    Who can use it: Claude Code on the web self-hosted environments.

    How to use it: When creating a new session, select `No repository` in the repository picker.

    Tip: Use this option for research or tool configuration tasks that don't require a repository to reduce unnecessary GitHub access.

    ✓ Verified by official sourceSource: Claude Code changelog
  • RetiredClaude Code

    subagentStatusLine removed; plugin agent setting deprecated

    The `subagentStatusLine` setting has been removed; subagent lines now use the default display, and the old key has no effect even if it remains in the file. The plugin's `agent` setting continues to work, but now triggers a warning at startup and in the output of `claude plugin list` and `validate`.

    Usage details

    Who can use it: Claude Code settings and plugin developers.

    How to use it: Remove the `subagentStatusLine` dependency; replace the use of `agent` in the plugin manifest with the newly supported configuration.

    Tip: Keep deprecation warnings visible in CI output and plan the migration before the next release.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Auto mode model can be selected with a managed setting for third-party providers

    Support for `CLAUDE_CODE_AUTO_MODE_MODEL` has been added to managed settings; it lets you specify the model used for Auto mode safety checks in Claude Desktop sessions using third-party providers. The setting takes effect when the relevant Claude Desktop version includes this Claude Code version.

    Usage details

    Who can use it: Managed Claude Desktop sessions and third-party providers.

    How to use it: In managed settings, set the `CLAUDE_CODE_AUTO_MODE_MODEL` environment variable to the model ID to use; deploy a Desktop version that includes 2.1.323.

    Tip: Verify that the safety-check model is actually available from the provider and that your organization’s policy allows it.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code SDK

    30-second progress signal added for long-running hooks

    Long-running hooks in `claude -p` and SDK sessions send a progress heartbeat every 30 seconds. This prevents a host such as Claude Desktop from assuming the turn is stuck and terminating it.

    Usage details

    Who can use it: Claude Code CLI `-p` mode and Agent SDK hosts; the host must include the new version.

    How to use it: Upgrade the CLI or SDK to 2.1.323; on custom hosts, handle heartbeat events as connection-liveness signals.

    Tip: Monitor heartbeats instead of simply increasing the host timeout; a real hang can be distinguished by a stopped heartbeat.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Numerous reliability fixes for VS Code, Windows, and accessibility workflows

    Fixes cover the sign-in screen, CRLF diffs, dictation, queued messages, and Fable warnings in VS Code, as well as PATH, deleted working directories, and long plugin paths on Windows. Screen readers, combining characters, terminal control sequences, key chords, and expanded dialog text are now handled more accurately.

    Usage details

    Who can use it: Claude Code CLI, VS Code/JetBrains panels, and Windows clients.

    How to use it: Upgrade to Claude Code 2.1.323; start a new session in VS Code, and reload the window if old panel state persists.

    Tip: Keeping project and plugin paths short on Windows is still useful; the fix does not guarantee that every tool will work with paths up to the 260-character limit.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code on the web

    Scheduled task and deletion warnings are now visible in web sessions

    A startup failure notice has been added for scheduled tasks that require a sandbox when the environment they run in does not have the necessary settings. The cloud session deletion dialog now clearly states that scheduled tasks running in that session will be paused.

    Usage details

    Who can use it: Claude Code on the web; cloud and self-hosted sessions.

    How to use it: Open the failure notice on the task card and match the sandbox requirement to the runtime environment; when deleting a session, move tasks to another session before confirming the task warning in the dialog.

    Tip: Test that the required sandbox is available in an environment before attaching scheduled tasks to a session.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code on the web

    Web repository picker and sandbox branch behavior are now safer

    If GitHub SSO organizations don't appear in the repository picker, the reason is shown along with a `Reconnect GitHub` link; the error for adding an inaccessible repository has also been replaced with a brief message and a reconnect button. When switching from an open folder to a web session with the VS Code sandbox enabled, VS Code no longer changes the branch automatically and explains why.

    Usage details

    Who can use it: Claude Code on the web and VS Code integration.

    How to use it: Select `Reconnect GitHub` on the Repository picker or Git providers page; choose the target branch yourself in the VS Code dialog.

    Tip: When making enterprise SSO changes, verify repository visibility on the GitHub connection page before checking Claude Code.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Desktop

    Web search and Web fetch can be managed separately from the Connectors menu

    Web search and Web fetch now appear as built-in tools under Connectors in the + menu and under Customize > Connectors. Each person can set them to Always ask or Blocked; `builtinToolPolicy` or `disabledBuiltinTools` set by device policy cannot be overridden.

    Usage details

    Who can use it: Claude Desktop v5.11492.0; only Web fetch is listed on Bedrock.

    How to use it: Open + > Connectors or Customize > Connectors; select Always ask or Blocked for Web search/Web fetch.

    Tip: Use Always ask for enterprise searches to keep external data access visible for every query.

    ✓ Verified by official sourceSource: Claude Desktop changelog
  • CapabilityClaude Code

    WebFetch, artifact database, and file upload issues fixed

    WebFetch no longer treats an empty model response as page content, no longer shifts empty cells in HTML tables, and uses less memory on deep pages. Page-by-page sorting guidance for large artifact collections, correct artifact previews, and rejection of filenames containing control characters have been added; cloud/Remote Control uploads are retried after a 507.

    Usage details

    Who can use it: Claude Code WebFetch, Artifact tool, cloud, and Remote Control.

    How to use it: Upgrade to 2.1.323; read large artifact collections page by page, and if a cloud upload fails, watch the retry status on the card.

    Tip: Provide WebFetch with the relevant sub-URL instead of a deep page to reduce both memory and context usage.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code Remote Control

    Remote Control sessions show a Try again button when the computer signs out

    A `Try again` button has been added to the error shown when the computer running the session signs out of Claude Code.

    Usage details

    Who can use it: Claude Code Remote Control.

    How to use it: Click `Try again` on the error card; if the issue persists, sign back in to Claude Code on the computer and restart Remote Control.

    Tip: Keep the Remote Control host up to date on machines that need a sign-in refresh.

    ✓ Verified by official sourceSource: Claude Code changelog

Google · Gemini

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

New in this edition · 16

  • AgentGemini Enterprise Agent Platform

    Agent Gateway for GKE entered preview in Gemini Enterprise Agent Platform

    Agent Gateway for GKE is now in Preview to centrally manage incoming, outgoing, and inter-agent traffic for agents running on GKE. Initial support is available on clusters running GKE 1.37.2-gke.1042000 or later in two regions.

    Usage details

    Who can use it: Versions `1.37.2-gke.1042000` or later available today in the GKE Rapid channel; `europe-west1` and `us-central1`. Other channels are expected to become available later.

    How to use it: Upgrade your GKE cluster to a supported version, then install the gateway and configure agent traffic using the Agent Gateway for GKE overview instructions.

    Tip: Before moving to production, try the Rapid channel version on a separate test cluster and verify the availability of other GKE channels.

  • APIGemini API

    Gemini 4 Hydrogen API adds xhigh thinking and 1 million-token long decode

    Gemini 4 Hydrogen offers `low`, `medium`, `high`, and the new `xhigh` thinking levels; the default is `high`. The combined 1,048,576-token window can go beyond the 65,536-token limit for a single request using `background` execution or continuation.

    Usage details

    Who can use it: The Gemini API’s GenerateContent and Interactions API endpoints, and the free and paid tiers of Google AI Studio. Long output requires an API client and the relevant long decode flow.

    How to use it: Select the `gemini-4-hydrogen` model; set `generation_config.thinking_level` to `low`, `medium`, `high`, or `xhigh`. For output exceeding 65,536 tokens, use `background=true` and an appropriate `max_output_tokens`, or implement client-side continuation.

    Tip: Start with `high` for most tasks; use `xhigh` only if measurements confirm a quality improvement, since it increases latency and token consumption.

  • AvailabilityGemini Enterprise

    Gemini 4 Hydrogen reaches GA in Gemini Enterprise apps and tools

    Gemini 4 Hydrogen is generally available in the Gemini Enterprise web and mobile apps, Workflow Builder, and AI developer tools. Organizations can manage model access; access in the web and mobile apps is off by default.

    Usage details

    Who can use it: `global`, `us`, and `eu` multi-regions; data residency and machine learning processing support for `us` and `eu`. It is enabled in AI developer tools for licensed users when the default Antigravity model list is active; organizations using a custom list must have an administrator select it.

    How to use it: As an administrator, enable Gemini 4 Hydrogen under Manage features > Model availability in the Gemini Enterprise web app. Select the model in Workflow Builder; also select Hydrogen in the custom AI developer tools model list.

    Tip: First verify your `us` or `eu` data residency requirements, then enable the model for a small group of users.

  • AvailabilityGoogle Search

    Gemini 4 Hydrogen is now available globally in Google Search

    With Gemini 4 Hydrogen, Google Search better understands the intent behind multi-part and misspelled queries. Answers provide more images, structured information, comparison tables, and related links; query fan-out scans a wider range of subtopics at the same time.

    Usage details

    Who can use it: Global rollout for all Google Search users; available in supported AI Mode languages.

    How to use it: Turn on AI Mode in Google Search and enter a natural-language query with multiple conditions. Review the images, tables, and expanded list of links in the results.

    Tip: State all the conditions in your query clearly in a single message to benefit from Hydrogen’s multi-part intent resolution.

  • ModelGemini

    Gemini 4 Hydrogen released for everyday and agent-based tasks

    Google announced Gemini 4 Hydrogen, a model focused on speed and efficiency. The model makes the Gemini 4 family available to a wider audience for coding, agent-based tasks, multimodal work, and everyday Gemini experiences.

    Usage details

    Who can use it: Global rollout; the free and paid tiers of Google AI Studio and Gemini API, the Gemini app, Google Search, Gemini for business, and Google Antigravity. In the Gemini app, free users access it through Auto model selection for eligible prompts; Antigravity access is available to paid Google AI Pro and Ultra subscribers.

    How to use it: In AI Studio, select Gemini 4 Hydrogen from the model picker; in the API, set the model value to `gemini-4-hydrogen`. In the Gemini app, select Hydrogen from the model picker or leave Auto model selection enabled.

    Tip: Start with Hydrogen for everyday coding and tool-using tasks; switch to Carbon or Argon for deeper architectural decisions.

  • AgentGoogle Antigravity

    Google Antigravity adds Orchestrator, which works across Hydrogen and Carbon

    Orchestrator has been added to the Google Antigravity model menu. Within the same session, Orchestrator dynamically delegates planning and difficult decisions to Carbon, and routine editing and testing to Hydrogen, balancing quality, quota, and latency.

    Usage details

    Who can use it: Gemini 4 Hydrogen is available to paid Google AI Pro and Ultra subscribers. Orchestrator (Carbon) and Orchestrator (Argon) are available on Ultra, while Orchestrator (Argon) is available on paid Pro; Gemini for business users can access both as well.

    How to use it: Open the model picker in Google Antigravity and select Orchestrator (Carbon) or Orchestrator (Argon). For programmatic routing, connect the same model roles to separate agents using Antigravity SDK.

    Tip: Choose Orchestrator when a single session involves both planning and execution; for entirely routine, fast tasks, using Hydrogen directly may provide lower latency.

    ✓ Verified by official sourceModels: Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 ArgonSource: Gemini 4 Hydrogen ve Orchestrator resmi Antigravity duyurusuSource: Gemini 4 Hydrogen resmi tanıtımı
  • APIBigQuery generative AI functions

    BigQuery generative AI functions gain Gemini 4 Hydrogen support

    BigQuery’s generative AI functions feature now supports the `gemini-4-hydrogen` model. Model support is marked as generally available.

    Usage details

    Who can use it: GA in Google Cloud projects using BigQuery generative AI functions.

    How to use it: Select `gemini-4-hydrogen` as the model ID in the relevant generative AI function configuration in BigQuery, then run the query.

    Tip: When switching models, compare query cost and token usage against your current model.

    ✓ Verified by official sourceModels: Gemini 4 HydrogenSource: Google Cloud 12 Kasım sürüm notları
  • Pricing / PlanGemini API

    Free tier and new API pricing announced for Gemini 4 Hydrogen

    Gemini 4 Hydrogen is available on a quota-limited free tier in Gemini API and AI Studio. The paid-tier price is 1,50 dolar per 1 million input tokens and 7,50 dolar per 1 million output tokens; cached input costs 0,075 dolar on implicit cache hits.

    Usage details

    Who can use it: Gemini API and Google AI Studio; the free tier is rate-limited, while the paid tier is subject to standard billing.

    How to use it: Call the `gemini-4-hydrogen` model in a paid project; manage your budget by tracking input, thinking, and output tokens together. Use implicit caching for long shared prefixes.

    Tip: To reduce costs, keep the thinking level at the lowest value that meets your requirements and take advantage of shared prompt prefixes exceeding 4,096 tokens.

  • FeatureGemini Enterprise

    Version history and restore are now available for Gemini Enterprise skills

    Skill owners can now view a skill’s version history on the Skills page, inspect instructions and files from previous versions in read-only mode, and restore an older version as the current version.

    Usage details

    Who can use it: Skills page for skill owners in Gemini Enterprise.

    How to use it: Open the Skills page, go to the version history for the relevant skill, review the instructions and files from an older version, and use Restore to make it the current version if needed.

    Tip: Before restoring, compare the files and instructions in the older version using the read-only view.

  • AgentGemini Managed Agents

    Antigravity agent default in Gemini Managed Agents is now Hydrogen

    Antigravity agent in Gemini Managed Agents has started using Gemini 4 Hydrogen as its default underlying model due to performance and reasoning improvements. The underlying model can be configured with `agent_config`.

    Usage details

    Who can use it: Gemini Managed Agents users in Gemini API and Google AI Studio.

    How to use it: Use `agent=antigravity-preview-09-2026` in an Interactions API call; configure the model settings in `agent_config` to change or pin the default.

    Tip: Measure quality, latency, and token costs before and after the transition to avoid being caught off guard by the default change in production.

    ✓ Verified by official sourceModels: Gemini 4 Hydrogen, Antigravity agentSource: AI Studio Gemini 4 Hydrogen yenilikleri
  • FeatureGemini in Chrome

    Gemini in Chrome highlights the workflow for summarizing web pages in the side panel

    Google’s official post dated November 12 describes using the Gemini icon in the Chrome toolbar to summarize the current web page in the side panel, compare products, and explain complex topics. The post does not specify a new plan or regional restriction.

    Usage details

    Who can use it: Chrome users who see the Gemini icon; the official post does not specify plan or regional coverage.

    How to use it: Click the Gemini icon in the Chrome toolbar to open the side panel; ask for a summary, comparison, or explanation without leaving the page.

    Tip: For long pages, first ask for a brief summary, then follow up about only the sections that affect your decision.

  • OtherGemini Notebook for business

    Gemini Notebook Enterprise renamed Gemini Notebook for business

    Google changed the name Gemini Notebook Enterprise to Gemini Notebook for business. Product capabilities remain unchanged, and API endpoints have not changed.

    Usage details

    Who can use it: Current Notebook users and API consumers in Gemini Enterprise documentation; the rename does not change product capabilities or endpoints.

    How to use it: Update the product name to Gemini Notebook for business in internal documentation, bookmarks, and user communications; continue using the existing endpoints in API calls.

    Tip: Don’t change integrations unnecessarily; this is only a naming change.

  • APIGemini Enterprise Agent Platform

    End-user `safety_identifier` labels for Gemini API calls are now GA

    API resellers, model aggregators, and multi-tenant applications can now include an anonymized `safety_identifier` value in requests. Google can associate a policy violation with a specific end user instead of applying it to the entire Cloud project.

    Usage details

    Who can use it: Global, multi-region, and regional endpoints for Gemini `generateContent` and `streamGenerateContent` calls; the feature is GA for API resellers and applications with external users and requires no additional fee or IAM permission.

    How to use it: Add `safety_identifier` to the `labels` field in the request body. The value must be an anonymous value 1–63 characters long, containing lowercase letters, numbers, underscores, or hyphens; do not use personal data such as an email address.

    Tip: Hashing a user ID with SHA-256 and truncating it to 63 characters provides a stable label without sending PII.

  • CapabilityGemini for Home

    Gemini for Home handles complex home commands more reliably

    For complex smart home requests involving multiple devices, Gemini for Home has started asking a brief clarifying question instead of guessing. The voice feedback flow also makes it easier to add details or change a message before sending it.

    Usage details

    Who can use it: Eligible Google Home/Nest setups using the Gemini for Home (Early Access) voice assistant.

    How to use it: Give a natural-language home command that affects multiple devices; if there’s ambiguity, Gemini will ask which device you mean. When providing feedback by voice, add details or change the text before sending it.

    Tip: Specify the room or device name at the start of the command to get a faster result without needing a clarification round.

  • CapabilityGemini for Home

    Gemini for Home starts showing visual answers on smart displays

    Gemini for Home (Early Access) can show relevant images alongside text answers on Nest Hub and compatible smart displays when visuals are useful. This means information is no longer limited to spoken answers.

    Usage details

    Who can use it: Compatible Google Nest Hub and smart displays included in Gemini for Home Early Access; Google may further limit rollout and regional availability.

    How to use it: Ask Gemini for Home a question on a compatible smart display included in Early Access; if a relevant image is available, it will appear on the screen with the answer.

    Tip: For questions about places, objects, or comparisons, ask for visual context; the on-screen answer is easier to scan.

  • FeatureGemini for Home

    Gemini for Home speeds up timer and Broadcast actions

    Gemini now sets, checks, or cancels timers, alarms, and media sleep timers faster; a rare issue when setting a sleep timer has also been fixed. Broadcast messages across the home now start playing faster on speakers and displays.

    Usage details

    Who can use it: Google Home/Nest speakers and displays that support Gemini for Home (Early Access).

    How to use it: Ask Gemini for Home to set, check, or cancel a timer, alarm, or media sleep timer; use Broadcast to send a message throughout the home.

    Tip: For media, request a sleep timer with a specific duration, and clearly say room names to reduce Broadcast delays across multiple rooms.

OpenAI · ChatGPT

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

New in this edition · 17

  • SettingCodex CLI

    Open and resumed chats retain their saved model provider

    Codex 0.168.0 ensures that open or resumed chats continue using their saved model provider when the default provider for new chats is changed. This prevents a provider change from unintentionally changing the backend for existing work.

    Usage details

    Who can use it: Codex CLI 0.168.0; TUI chats using multiple model providers.

    How to use it: Update Codex to 0.168.0; change the default for new chats in Sessions → Model providers, then open an older chat to verify that it resumes with its saved provider.

    Tip: Start new work before switching providers; check the chat history to see which provider is being used for ongoing long-running tasks.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Responses interrupted by tool calls are displayed as a single Markdown message

    Codex 0.168.0 renders TUI responses interrupted by tool calls as a single Markdown message and restores reasoning summaries when replaying a conversation. This improves the readability and resumability of long sessions involving tools.

    Usage details

    Who can use it: Codex CLI 0.168.0; users who view and replay TUI conversations.

    How to use it: Update Codex CLI to 0.168.0; open or resume a chat interrupted by a tool call, then use the single, continuous Markdown response and restored reasoning summary.

    Tip: Before sharing a long session, reopen the chat and check that the rendered message and reasoning summary appear complete.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Steering messages now reach running turns during final transcript recording

    Codex 0.168.0 fixes an issue where steering messages sent during final transcript recording or after returning to a long-running conversation did not reach the running turn. Users can now steer ongoing work even at later stages.

    Usage details

    Who can use it: Codex CLI 0.168.0; long-running or reactivated running TUI turns.

    How to use it: Update Codex to 0.168.0; while a long-running turn is active, send a steering message, especially during final recording or when returning to a conversation, and wait for the turn to process the new instruction.

    Tip: Steer a running turn with small messages instead of canceling it; still verify critical changes in an explicit next turn.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Codex TUI can open a resizable side conversation

    Codex 0.168.0 offers a second conversation pane beside the main chat with the experimental tui_workspace feature. The panes can be resized by dragging, and the focused pane can be enlarged with a shortcut.

    Usage details

    Who can use it: Codex CLI 0.168.0; TUI users who enable the feature in the configuration.

    How to use it: Enable features.tui_workspace = true in config.toml; restart Codex, drag the pane divider, or press leader z to zoom the focused pane.

    Tip: Keep research or debugging context in the side pane without disrupting the main chat, and compare both workflows on the same screen.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • FeatureCodex CLI

    Codex can generate animated GIFs with a configuration flag

    Codex 0.168.0 adds the image_generation_animation feature, which enables creating animated GIFs from image generation or editing. Optional controls are available for frame count and timing.

    Usage details

    Who can use it: Codex CLI 0.168.0; users who enable the animation feature and have access to a compatible image model.

    How to use it: Set features.image_generation_animation = true in config.toml; ask Codex to create or edit an animated GIF and, if supported, specify the frame count and timing.

    Tip: Keep the frame count and duration short for quick previews; separately check the balance between detail and file size in the final GIF.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • SettingCodex CLI

    Codex adds a provider catalog and a default selection for new chats

    Codex 0.168.0 lets you view available model providers and select a default provider for new chats in Sessions → Model providers. Open and resumed chats retain their saved providers.

    Usage details

    Who can use it: Codex CLI 0.168.0; users with provider configurations.

    How to use it: Update Codex to 0.168.0; open the Sessions → Model providers menu, select a provider, and save it as the default for new chats.

    Tip: When comparing different providers, change the default for new chats, but continue existing long-running work with its saved provider.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Recording attachments are supported with audio-capable models

    Codex 0.168.0 lets you attach recordings to a chat when using an audio-capable model. Recordings rejected after approval are skipped in subsequent turns, and clip duration is taken into account to keep short recordings in context.

    Usage details

    Who can use it: Codex CLI 0.168.0 and an audio-capable model; subject to model and account access.

    How to use it: Update Codex to 0.168.0, select an audio-capable model, and attach a recording to the prompt; use the existing context for short recordings rather than attaching them again in a subsequent turn.

    Tip: Use short, focused recordings; splitting long recordings makes it easier for the model to keep the relevant section in context.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Local history and resume selectors are filtered by ChatGPT account

    Codex 0.168.0 filters newly created threads in local TUI history and resume selectors by the signed-in ChatGPT account. This reduces the chance of seeing the wrong chats when working across accounts on the same computer.

    Usage details

    Who can use it: Codex CLI 0.168.0; local TUI users signed in with a ChatGPT account.

    How to use it: Update Codex CLI to 0.168.0, sign in with your ChatGPT account, and open the history or resume selector; the list will show threads belonging to your account.

    Tip: If you use multiple accounts, switch accounts before moving to your work account and check that you create the new thread under the correct account.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Fuzzy file search ranks filenames higher

    Codex 0.168.0 gives filename matches more weight in fuzzy file search. This makes it less likely that results from matching parent directories will crowd out the correct filenames.

    Usage details

    Who can use it: Codex CLI 0.168.0; developers using TUI file search.

    How to use it: Update Codex CLI to 0.168.0 and use file search as usual; filename matches should appear ahead of directory matches in the results.

    Tip: In large repositories, include a distinctive part of the filename in the search term; the new weighting brings the right file to the top faster.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • SettingCodex CLI

    Codex terminal notifications and approval previews are expanded

    Codex 0.168.0 supports terminal notifications with the tui.notification_method = "osc777" setting. Approval notifications now show longer command previews and distinguish network access requests from terminal input.

    Usage details

    Who can use it: Codex CLI 0.168.0; terminal environments that support OSC 777 notifications.

    How to use it: Set tui.notification_method = "osc777" in config.toml; restart Codex and use the distinction between command and network requests in approval notifications.

    Tip: Enable OSC 777 for long-running tasks and monitor approval notifications on your desktop; review network access prompts separately before running commands.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • CapabilityCodex CLI

    Visual previews and Pets support for Ghostty and Kitty in tmux

    Codex 0.168.0 supports visual previews and Pets display when using Ghostty and Kitty in tmux full-screen mode. Terminal passthrough permission must be enabled for this to work.

    Usage details

    Who can use it: Codex CLI 0.168.0; Ghostty or Kitty environments using tmux and the allow-passthrough all setting.

    How to use it: Enable allow-passthrough all for tmux in the Ghostty or Kitty terminal configuration, run Codex full-screen, and open visual or Pets output.

    Tip: In corporate or shared terminals, enable passthrough only in tmux sessions you trust.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • Plugin / IntegrationCodex CLI

    Idle local MCP servers preserve file access on Windows

    Codex 0.168.0 fixes an issue where idle local MCP servers on Windows blocked filesystem I/O elsewhere in Codex. File operations now continue more reliably with a local MCP integration enabled.

    Usage details

    Who can use it: Codex CLI 0.168.0; users running local MCP servers on Windows.

    How to use it: Update Codex CLI to 0.168.0, configure your local MCP server, and run file operations as usual while the server is idle.

    Tip: For teams using MCP on Windows, after updating, verify both an MCP call and a concurrent file scan in a small test repository.

    ✓ Verified by official sourceSource: Codex CLI 0.168.0 yayın notları
  • APIOpenAI Go SDK

    Go SDK enables reasoning input transformations in Responses outputs

    OpenAI Go SDK v3.99.0 makes reasoning input transformations information in Responses outputs parsable and accessible. This makes it easier to use the new API feature with typed responses in Go.

    Usage details

    Who can use it: Developers using OpenAI Go SDK v3.99.0 and Responses API integrations.

    How to use it: Update the SDK with go get github.com/openai/openai-go/v3@v3.99.0 and continue using your existing client for Responses API calls; inspect the reasoning input transformation fields parsed by the SDK in the response.

    Tip: Before converting response objects to raw JSON, use the v3.99.0 types to get compile-time checks for transformation information.

    ✓ Verified by official sourceSource: openai-go v3.99.0 sürümleri
  • APIOpenAI Node SDK

    Node SDK adds types for Responses reasoning input transformations

    OpenAI Node SDK v7.56.0 provides TypeScript types for reasoning input transformations information in Responses outputs. This makes the new response structure safer to use in JavaScript and TypeScript applications.

    Usage details

    Who can use it: JavaScript and TypeScript Responses API applications using OpenAI Node SDK v7.56.0.

    How to use it: Update the package with npm install openai@7.56.0; use the reasoning input transformation fields provided by the SDK in the response to the client.responses.create call.

    Tip: If you use TypeScript, preserve the response types directly and take advantage of compiler checks without casting transformation fields to any.

    ✓ Verified by official sourceSource: openai-node v7.56.0 sürümleri
  • APIOpenAI Python SDK

    Python SDK adds typed reasoning input transformations in Responses

    OpenAI Python SDK v3.50.0 exposes reasoning input transformations information in Responses outputs. Developers can access these new transformations through the Python SDK’s typed objects.

    Usage details

    Who can use it: Responses API developers using OpenAI Python SDK v3.50.0.

    How to use it: Update the SDK with pip install --upgrade openai==3.50.0; use the reasoning input transformation information in the typed object returned by your Responses API call.

    Tip: After upgrading the SDK, validate the response schema in a test environment; retain fallback error handling, especially for unknown or empty responses.

    ✓ Verified by official sourceSource: openai-python v3.50.0 sürümleri
  • APIOpenAI Ruby SDK

    Ruby SDK exposes Responses reasoning input transformations

    OpenAI Ruby SDK v0.123.0 makes reasoning input transformations information available in Responses outputs. Ruby clients can process the new transformations through the SDK’s response objects.

    Usage details

    Who can use it: Responses API developers using OpenAI Ruby SDK v0.123.0.

    How to use it: Update the gem with gem install openai -v 0.123.0; read the reasoning input transformation information from the object returned by your Responses API call.

    Tip: Before deploying to production, test existing serialization and nil behavior with the new gem; retain a compatibility path for older response formats.

    ✓ Verified by official sourceSource: openai-ruby v0.123.0 sürümleri
  • Pricing / PlanChatGPT Plus

    Free year of Plus continues for U.S. service members, veterans, and retirees

    OpenAI’s November 11 update reports that more than 20,000 transitioning service members and veterans have used the one-year ChatGPT Plus offer. The offer provides one year of free Plus access to eligible U.S. service members, veterans, and retirees after verification with ID.me.

    Usage details

    Who can use it: Only verified U.S. service members, veterans, and retirees who are within 12 months of leaving active duty or have left within the past 12 months; eligible personal ChatGPT accounts. Enterprise workspace accounts and active Plus subscriptions through mobile app stores are not eligible.

    How to use it: Sign in with your personal account at chatgpt.com/veterans-claim; select Verify with ID.me and complete military verification through ID.me; enter your service end date if required, then select Claim offer after Eligibility confirmed.

    Tip: The offer can be used once and automatically renews at the standard monthly price after one year; if you don’t want to renew, set a reminder to cancel before the promotion ends.

Meta · Meta AI

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

New in this edition · 4

  • SettingWhatsApp Cloud API

    WhatsApp adds the `call_permission_mode` setting for quality-based business calls

    The Cloud API added an optional `call_permission_mode` setting that lets businesses opt in to or out of quality-based business-initiated calls. The active mode is read from the Settings GET response, and mode changes are reported through the `account_settings_update` webhook.

    Usage details

    Who can use it: Eligible business phone numbers with the Calling API enabled and covered by quality-based business-initiated calling; WhatsApp Cloud API and Webhooks surfaces.

    How to use it: Set the opt-in preference by sending the `call_permission_mode` field in a `POST /<PHONE_NUMBER_ID>/settings` call; read the active mode with `GET /<PHONE_NUMBER_ID>/settings` and subscribe to the `account_settings_update` webhook field.

    Tip: Don’t treat local configuration as the sole source of truth; sync the active mode from the Settings GET response and changes from the webhook.

  • FeatureWhatsApp Embedded Signup

    Business display name screen added to Embedded Signup v4/v5 previews

    Meta added a screen to the v4-public-preview and v5-public-preview flows that asks for the phone number’s display name after business assets are selected or created. The entered name is verified as the name WhatsApp users will see instead of the phone number; rollout of the screen to stable v4 and v5 versions will follow later.

    Usage details

    Who can use it: v4-public-preview and v5-public-preview for all business customers; rollout to stable v4 and v5 versions will follow later.

    How to use it: Use the preview flow by setting `extras.version` to `v4-public-preview` or `v5-public-preview` in the Embedded Signup launch configuration, and include the display name screen in integration tests.

    Tip: Verify in a test environment that the Next button is disabled for unapproved names and that the screen is skipped for numbers that already have an approved display name.

  • AvailabilityWhatsApp Cloud API

    Manual onboarding removed for WhatsApp Pix with redirect

    The official changelog dated November 11 announced the removal of the manual onboarding requirement for Pix with redirect. The bank redirect flow is selected by setting `flow_type` to `APPSWITCH` in the payment message; because the details page still includes the old manual onboarding notice, verify the rollout status for your account.

    Usage details

    Who can use it: Pix-enabled WhatsApp Cloud API businesses in Brazil; the redirect flow is available in one-to-one chats for eligible Android and iOS users, while on other surfaces the same message falls back to a copy-and-paste flow.

    How to use it: Add `flow_type: "APPSWITCH"` to the `payment_settings[].pix_dynamic_code` object in the Order details message; also send the dynamic Pix code and the `merchant_name`, `key`, and `key_type` fields. According to the November 11 changelog, manual onboarding through the intake form is not required.

    Tip: Always include a valid dynamic Pix code in the message so users can pay by copying and pasting if their bank or client doesn’t support the redirect.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 12

  • FeatureKimi Code Desktop

    Change the primary working directory in Desktop projects

    With v1.3.4, you can change the primary directory of a multi-directory Desktop project. This makes it easier to choose the project root again and correct the starting point of the file tree.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; projects with multiple working directories.

    How to use it: Edit the project in Project or workspace management and choose the working directory you want as the primary directory; then check the file tree root selection.

    Tip: Keeping the repo root as the primary directory and dependent apps or sample folders as source folders helps keep file and Git context more organized.

  • SettingKimi Code Desktop

    Choose separate UI and code fonts in Kimi Code Desktop

    With v1.3.4, UI, transcript, code, and terminal fonts can be customized independently; bundled Kimi Code fonts are also available.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; desktop users who upgrade to the November 12 release.

    How to use it: Open the Settings panel; change the UI, transcript, code, and terminal font fields independently, or choose one of the bundled Kimi Code fonts.

    Tip: Choose a readable monospace font for code and the terminal and a more comfortable reading font for transcripts to reduce eye strain during long sessions.

  • CapabilityKimi Code CLI ve Kimi Code Desktop

    Retry, partial responses, and background tasks work more reliably

    The retry indicator that gets stuck on a new attempt is cleared, partial text from failed attempts is preserved, and foreground tasks publish live status updates. Tool calls assigned to the background no longer hold the scheduler lock unnecessarily, so subsequent tasks can start in the same turn.

    Usage details

    Who can use it: Kimi Code CLI and Desktop 2.6.0 canary; users who retry after provider errors and run parallel background tool calls.

    How to use it: Use the retry button after a provider error; check the `Requesting…` status and the previous partial text on the next attempt. You can run tasks in parallel by starting multiple background Bash tasks in the same turn.

    Tip: Use retry for temporary provider errors instead of resubmitting the request from scratch; this preserves the previous attempt’s partial output.

  • AgentKimi Code CLI

    Tower adds the human-approved `artifact` task type

    Tower now supports the `artifact` mission type for reports, build outputs, and similar deliverables. This type requires the main deliverable files to be committed and explicit human acceptance before merging.

    Usage details

    Who can use it: Kimi Code CLI v2.6.0 documentation and the November 12 canary release; teams using the experimental Tower mode.

    How to use it: Enable Tower with `KIMI_CODE_EXPERIMENTAL_TOWER=1 kimi` or with `[experimental] tower = true` in config.toml; after `/tower on`, start the task as an `artifact` mission kind, specifying key files and human acceptance criteria.

    Tip: Clearly specify the files to be delivered and the acceptance criteria before starting an artifact task, so Tower doesn’t try to merge an incorrect intermediate output.

  • FeatureKimi Code CLI

    Experimental next-prompt suggestions arrive for TUI

    In the interactive TUI, an experimental next-prompt suggestion appears as ghost text when a turn is complete, the input field is empty, and the session is idle. Press Tab or the right arrow to accept and edit the suggestion; it is not sent unless you press Enter.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 canary; interactive TUI only, and the experimental feature is off by default.

    How to use it: Enable prompt prediction in the `/experiments` panel or launch with `KIMI_CODE_EXPERIMENTAL_PROMPT_PREDICTION=1`; accept the suggestion with Tab or the right arrow, edit it, and press Enter to send.

    Tip: Verify the suggestion before sending it directly; narrow down next steps especially when they involve changing files or require tool calls.

  • FeatureKimi Code Desktop

    Long pasted text is automatically folded in Composer

    With v1.3.4, long text pasted into Composer is automatically folded into a snippet view. Users can expand it to review or edit the full text before sending.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; desktop users who paste long text.

    How to use it: Paste long text into Composer; click the resulting folded snippet to view or edit the full text, then send it.

    Tip: Before sending large logs or code blocks, expand the snippet and remove unnecessary sections.

  • SettingKimi Code Desktop

    Local file tabs open muted by default

    When local files are opened from the built-in browser address bar, new tabs start muted and block audio autoplay. Per-site or global audio controls are also available.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; desktop users who use the built-in browser.

    How to use it: Open a local file from the address bar in Built-in Browser; use the tab’s audio control or the browser’s audio settings to allow audio for the site or all tabs.

    Tip: Keep the default muted behavior to prevent unintended audio playback when testing HTML previews.

  • CapabilityKimi Code CLI

    `/btw` forks now retain the main Agent’s baseline tool set

    Forks with the same profile now inherit the main Agent’s exact baseline tool list. This reduces parent-child tool mismatches if the tool profile changes during a session and preserves the prompt-cache prefix.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 canary; fork and side-question flows with the same profile, such as `/btw`.

    How to use it: After changing the tool profile or MCP tools in a session, open a side question with `/btw`; no additional settings are needed, and the fork uses the main session’s tool set.

    Tip: Using `/btw` questions in long-context sessions may reduce cache loss caused by resending tool definitions.

  • SettingKimi Code CLI

    Set TUI animation speed with `KIMI_CODE_TUI_ANIMATION_FPS`

    The new environment variable sets the Fullscreen TUI animation clock speed between 1 and 60 FPS. If the variable is not set, the current default behavior is preserved; invalid values are ignored.

    Usage details

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

    How to use it: Before starting the CLI, set a value between 1 and 60, such as `export KIMI_CODE_TUI_ANIMATION_FPS=30`, then run the `kimi` command.

    Tip: Reduce CPU usage on low-powered machines with 15–30 FPS; try 60 FPS for a smoother display.

  • FeatureKimi Code Desktop

    Move Desktop terminal tabs between the bottom and right panels

    v1.3.4 lets you move terminal tabs between the bottom and right panel positions. This lets you place the terminal to suit your workspace in narrow or wide window layouts.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; desktop users who use the terminal panel.

    How to use it: Open the terminal panel; use the terminal tab’s move menu to send the tab to the bottom or right panel.

    Tip: Choose the right panel when comparing code, and the bottom panel when you need more horizontal space; adjust the layout to suit the task.

    ✓ Verified by official sourceSource: Kimi Code Docs — What's New, v1.3.4
  • CapabilityKimi Code Desktop

    Pixel zoom and color copying added to screenshots

    v1.3.4 adds pixel zoom and color value copying tools to the screenshot workflow. This makes it easier to inspect small UI details and color differences.

    Usage details

    Who can use it: Kimi Code Desktop v1.3.4; users of the built-in screenshot tool.

    How to use it: In Composer, use + → Screenshot to capture an area; use the pixel zoom and color copying controls in the image tools.

    Tip: For UI bug reports, zoom in on the affected pixel area and send the color value directly to the designer or Agent.

  • CapabilityKimi Code CLI

    Improved sweep animation for Fullscreen running titles

    The light sweep in running transcript group and Thinking titles now moves toward the text color in both dark and light themes. The sweep starts after 600 ms and then repeats about every four seconds.

    Usage details

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

    How to use it: Select `/settings` → TUI mode → `fullscreen`; the animation appears in running group and Thinking titles without requiring an additional setting.

    Tip: If the animation is distracting, switch back to regular TUI or lower the FPS setting.

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 19

  • AvailabilityGrok Bot

    Grok Bot 1.0 opens to team use after beta

    xAI’s official X account announced Grok Bot 1.0 on November 11 and said Bots are shareable, more collaborative teammates. The official changelog entry from the same day says 1.0.1 also includes the changes from 1.0.0.

    Usage details

    Who can use it: Team Bots are available on Team and Enterprise plans; the meeting feature is also available on Team, Enterprise, and Ultra plans. Although the x.ai news page shows the announcement as dated November 2, the official X post is dated November 11.

    How to use it: Update the Grok Bot app; create a Bot with + > Create a Bot, or add a shared Bot with + > Browse shared Bots.

    Tip: Before sharing, review the Bot’s connected apps, secrets, skills, and files; teammates may use them in their chats.

  • Plugin / IntegrationGrok Bot

    Grok Bot can join Google Meet, Zoom, and Microsoft Teams meetings

    The Bot can join a Google Meet, Zoom, or Microsoft Teams meeting from a meeting link, take notes, answer questions in the meeting chat, and send a summary and action items at the end. A Bot that tracks its calendar can also join scheduled meetings automatically.

    Usage details

    Who can use it: Team, Enterprise, and Ultra plans; in Enterprise, an administrator must enable the Meetings feature.

    How to use it: Paste the Google Meet, Zoom, or Microsoft Teams link into the Bot chat and ask it to join and take notes; approve the confirmation card and have the meeting host let the Bot in. You can manage scheduled meetings under Upcoming in Bot details.

    Tip: Get the summary and action items directly in the Bot chat at the end of the meeting; longer reports are written to a Page if Pages is in use.

  • Plugin / IntegrationGrok Bot

    Team Bots can connect to Microsoft Teams

    Published Team Bots can now be used in Microsoft Teams channels, direct chats, and group chats. The Bot acts on behalf of the person making the request, can work with files, and asks for approval before risky actions.

    Usage details

    Who can use it: Published Team Bots on Team and Enterprise plans; the team administrator must enable Microsoft Teams, and the Microsoft Teams administrator must approve each Bot.

    How to use it: In the Grok Bot team dashboard, select Shared Bots > Enable Microsoft Teams. The Bot owner or Manager should go to Bot details > Channels > Connect Microsoft Teams and sign in with a Microsoft account; then approve it under Teams admin center > Teams apps > Manage apps > Pending approval.

    Tip: Setup can take up to an hour after approval; before communicating, check that the Bot appears in Teams and that your account is connected under Settings > General > Team Bots.

  • Plugin / IntegrationGrok Bot

    Team Bots can use shared app logins

    The Bot owner or Manager can share their own app login as a shared connector, so team members don’t need to connect separate accounts for the same app. This feature requires the team administrator to enable Allow Shared Logins.

    Usage details

    Who can use it: Teams using Team Bots; Allow Shared Logins is off by default in the team dashboard, and anyone accessing the shared login acts through the Manager’s account.

    How to use it: The team administrator should enable Sharing > Allow Shared Logins on the Grok Bot page. The Bot owner or Manager should select the app under Bot details > Setup > Service accounts > Shared connectors and use Share your account.

    Tip: Use only a limited-permission, non-sensitive service account for sharing; changes made through the shared login take place through the Manager’s account.

  • SettingGrok Bot

    0.5x–2x speed range and voice cards added for Bot voice

    A Speed slider ranging from 0.5x to 2x and a separate selection card for each voice have been added to Bot’s Voice menu. The same controls are also available in voice call settings.

    Usage details

    Who can use it: Users of Grok Bot’s voice chat and calling flows; the changelog does not specify plan or country restrictions.

    How to use it: Open the Voice menu in the Bot profile or chat settings; select a voice from its card and adjust the Speed slider from 0.5x–2x. You can also change it in call settings during a call.

    Tip: Try 1.25x–1.5x for listening to meeting notes, and 0.75x–1x for detailed content or content in a foreign language.

  • SettingGrok Bot

    Review learnings approval flow added for Team Bot learnings

    Team memory or skill changes taught to a Team Bot by teammates are no longer saved directly. Changes wait under Review learnings, where the owner or Manager can finalize each suggestion with Approve or Reject.

    Usage details

    Who can use it: Team Bots; applies to owner and Manager roles, and personal learnings are not added to team memory.

    How to use it: Open Review learnings under Bot details > Suggested; review what the suggestion does and select Approve or Reject. It can also be managed on mobile via Bot profile > Learnings.

    Tip: Approve only suggestions that genuinely change the team standard for shared memory; don’t save personal preferences as team learnings.

  • CapabilityGrok Bot

    Team Bots can read team comments on Pages

    Comments that teammates leave on Team Bot’s Pages content now reach the Bot. The Bot can take feedback on a shared Page into account in subsequent work.

    Usage details

    Who can use it: Teams using Team Bot and Pages; whether comments reach the Bot depends on team sharing permissions.

    How to use it: Open a Page created by Team Bot, add a comment, and wait for it to be processed in the relevant Team Bot chat.

    Tip: Clearly state the expected output, scope, and deadline in your comment so the Bot can make the requested change.

  • Plugin / IntegrationGrok Bot

    Custom MCP connectors in Team Bots shown as shared connections

    Custom MCP connectors used in a Team Bot are now shown as shared with everyone on the team. The Share my account option does not appear for these connections, distinguishing shared MCP configuration from personal account sharing.

    Usage details

    Who can use it: Team Bots using custom MCP connectors; team policies and MCP allowlist permissions apply.

    How to use it: Check the custom MCP connector status under Team Bot details > Setup > Service accounts or Connect Apps; shared status is applied automatically to connections permitted by the team administrator.

    Tip: Don’t embed personal credentials in custom MCP configuration; use separate, limited-permission credentials for shared team connections.

  • SettingGrok Bot

    Disable Interactivity setting added for meeting note-taking

    A Bot scheduled for a meeting continues listening and taking notes when Disable Interactivity is selected, but does not respond in the meeting chat. The meeting row is shown as Muted in this case.

    Usage details

    Who can use it: Team, Enterprise, and Ultra plans using the Grok Bot Meetings feature; Meetings must be enabled by an administrator in Enterprise.

    How to use it: In Bot details > Upcoming, open the menu for the relevant meeting and select Disable Interactivity. To have it respond again, select Enable Interactivity; if the Bot has joined or is in the meeting, it must be removed and added again.

    Tip: Turn this setting on if you only need a transcript; leave it off if you want immediate answers to questions in the meeting chat.

    ✓ Verified by official sourceSource: Grok Bot’a toplantı ekleme
  • SettingGrok Bot

    Admins can set Team Bot sharing to Everyone or Admins only

    Team administrators can now limit who can share Bots with the Bot sharing setting. When Admins only is selected, members’ Share button is disabled and the message Only team admins can share Bots is displayed; Bots shared previously remain shared.

    Usage details

    Who can use it: Team and Enterprise teams; the default option is Everyone, with Admins only as the alternative.

    How to use it: In the Grok Bot team dashboard, open Shared Bots > Sharing > Bot sharing and select Everyone or Admins only.

    Tip: For Team Bots connected to production or sensitive data, set up a review gate with Admins only first; note that existing shares are not removed automatically.

  • AvailabilityGrok Bot

    Brazilian Portuguese localization made more natural

    Brazilian Portuguese text in the app has been improved to read more naturally. This is not the addition of a new language, but a quality update to the existing user-facing localization.

    Usage details

    Who can use it: Grok Bot users with the Brazilian Portuguese interface.

    How to use it: Select Brazilian Portuguese under Language in Grok Bot settings, or reopen the app if your device language is Brazilian Portuguese.

    Tip: If you encounter a localization issue, send the example text and screen context through Send Feedback.

  • CapabilityGrok Bot

    New Bot setup that takes longer now shows Still creating

    If setup for a new Bot takes longer than expected, the app now shows Still creating instead of an error. Text the user enters during this time is also preserved in the input box.

    Usage details

    Who can use it: New Grok Bot creation flow; no plan or platform restrictions specified.

    How to use it: Create a Bot with + > Create a Bot; if setup takes a while, wait for the Still creating status and continue by sending the text in the input box.

    Tip: Don’t resend the same request repeatedly while setup is in progress; when the status updates, send the preserved text once.

  • FeatureGrok Bot

    Typing -> in the message box automatically converts it to an arrow

    When you type -> in the message box, the input now converts it to an arrow. This offers a small typing convenience for short text describing a flow, relationship, or direction.

    Usage details

    Who can use it: Grok Bot message box; no plan or platform restrictions are specified.

    How to use it: Type -> in the message box; the app automatically converts the characters to an arrow.

    Tip: If you don’t want the automatic conversion, try writing the expression differently instead of copying and pasting the arrow.

  • FeatureGrok Bot

    Negative feedback submission no longer asks why

    When you give a response a thumbs down, the flow now ends with a thank-you message without asking why. This makes it possible to submit negative feedback in fewer steps.

    Usage details

    Who can use it: Grok Bot chat responses; no plan or platform restrictions are specified.

    How to use it: Tap or click the thumbs down button below the response; the feedback is submitted without opening an additional reason form.

    Tip: If you want to explain a specific issue, use the Send Feedback flow instead of thumbs down.

  • CapabilityGrok Bot

    No standalone done message when there’s no result to report

    When there’s nothing to report, the Bot no longer sends a separate message just to say the task is done. This reduces unnecessary notifications, especially for routine and background work.

    Usage details

    Who can use it: Grok Bot chats and background tasks; no plan or platform restrictions specified.

    How to use it: No additional settings are needed; a Bot turn that produces no result now stays silent instead of sending a status message like done.

    Tip: Check the routine history or relevant Bot chat to see whether a routine actually produces output.

  • SettingGrok Bot

    Optional device diagnostics added to Send Feedback

    The Send Feedback flow now includes the Include additional device diagnostics option. It is off by default, and additional device diagnostic data is sent only if the user selects it.

    Usage details

    Who can use it: Users of the Send Feedback flow in the Grok Bot app; the option is off by default.

    How to use it: Open Send Feedback and check the Include additional device diagnostics box only if you think it is necessary for support review.

    Tip: For privacy, leave the box unchecked by default; enable it only for issues where the support team needs device context.

  • FeatureGrok Bot

    Check price again option added to expired order cards

    An expired order card now shows a Check price again option. Selecting it asks the Bot for a new order card with current pricing and stock information.

    Usage details

    Who can use it: Grok Bot shopping flows; availability may vary depending on the store and connected connector support.

    How to use it: Tap Check price again on the expired order card and review the updated order card that is generated.

    Tip: Before paying, check the current total, stock status, and delivery information again.

  • FeatureGrok Bot

    Wide tables in responses expand into the available space

    Wide tables in responses now expand into the available space beside the message. This reduces the need for horizontal scrolling and cramped cells.

    Usage details

    Who can use it: Chat interface covered by the Grok Bot 1.0.1/1.0.0 changelog; no plan or platform restrictions are specified.

    How to use it: No action is required; open a response containing a wide table and view it in the space beside the message.

    Tip: For results with many columns, use the desktop app to view the table at full width.

  • SettingGrok Bot

    Connectors disabled by an administrator are clearly labeled

    When a team administrator disables a connector, its connection button now shows the text Disabled by team admin. This makes it clear that the issue stems from team policy, not the user’s account.

    Usage details

    Who can use it: When administrator connector policies are applied in Team and Enterprise teams.

    How to use it: Check the status of the relevant connector in Connect Apps; if Disabled by team admin appears, ask the team administrator for permission or wait for them to change the allowlist policy.

    Tip: In this case, check the team administrator’s connector policy first instead of trying to sign in again.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 48

  • OtherAI-Infra-Guard

    24 new AI product CVE advisories added from NVD

    Another 24 AI product security advisories from NVD have been added to the AI-Infra-Guard dataset. This is a separate data update following the v4.6.10 package on November 12.

    Usage details

    Who can use it: AI-Infra-Guard’s NVD-based advisory data; scans using the latest data package.

    How to use it: Update the data package or install the latest AI-Infra-Guard version and rerun the advisory scan.

    Tip: Merge the new advisories with your existing inventory and separately verify affected versions.

  • CapabilityAI-Infra-Guard

    Fingerprint rule added for One-API

    A new rule that recognizes the One-API product has been added to AI-Infra-Guard’s fingerprint checks. This expands the automatic classification of One-API traces and related installations.

    Usage details

    Who can use it: AI-Infra-Guard fingerprint scans and AI infrastructure using One-API.

    How to use it: Run the latest fingerprint scan; if One-API is installed, the new rule will be applied automatically in the results.

    Tip: Verify the fingerprint result along with the version and installation path; product detection alone is not evidence of a security vulnerability.

  • CapabilityAI-Infra-Guard

    Skill-Scan retries on transient LLM errors

    Skill-Scan retries with increasing delays on transient 300-second, connection, 429, and 5xx LLM errors. These errors can be recovered from before they turn into permanent scan failures.

    Usage details

    Who can use it: AI-Infra-Guard Skill-Scan; LLM providers that return transient errors.

    How to use it: Run the latest Skill-Scan as usual; built-in retries will kick in for transient errors.

    Tip: Reduce concurrent scans when providers impose rate limits to prevent retries from overlapping.

  • SettingLoopForge

    The SOP role can now select any built-in CLI

    In LoopForge’s web layer, the SOP role can select any available built-in CLI, rather than being limited to a previously narrowed set of options. This expands role-specific executor configuration.

    Usage details

    Who can use it: LoopForge’s SOP role configuration and built-in CLI catalog.

    How to use it: Open the SOP role settings, select an available built-in tool from the CLI selector, and save the workflow.

    Tip: Select a CLI with only the permissions each role needs.

  • FeatureLoopForge

    The run workspace can be edited while an agent is running

    LoopForge allows run workspace content to be edited while an agent is running. Content hashes and conflict checks help reduce accidental overwrites of the agent’s work.

    Usage details

    Who can use it: LoopForge run workspaces running agents.

    How to use it: Open the workspace editor on the active run page, save your changes, and merge with the latest version if a content conflict warning appears.

    Tip: Check the run status before making manual edits while the agent is modifying the same file.

  • FeatureMemHub

    Corrupted knowledge graph quarantined and rebuilt automatically

    MemHub 1.0.7 opens a corrupted or stuck Kuzu graph in a separate probe process with limited time and memory. If that fails, it preserves the graph.quarantined file, starts Core without a graph, and rebuilds the graph per user. The application and status surfaces show the rebuild status.

    Usage details

    Who can use it: MemHub 1.0.7; Linux x64/arm64 Runtime and macOS arm64 App/Runtime.

    How to use it: Upgrade MemHub to 1.0.7; automatic quarantine and rebuild are applied when the graph is corrupted. Advanced users can adjust ONE_MEMORY_GRAPH_PROBE_TIMEOUT_S and ONE_MEMORY_GRAPH_PROBE_MAX_RSS_MB.

    Tip: Keep quarantine files for diagnostics before deleting them; check the memory count after rebuilding.

    ✓ Verified by official sourceSource: MemHub 1.0.7 resmi sürüm commit’i
  • FeatureBrowserSkill

    Plan flows are now available in BrowserSkill CLI

    BrowserSkill CLI adds the bsk plan flow, which runs agreed-upon actions in sequence and stops at the first incomplete step. Plans can also be used in MCP and DSH.

    Usage details

    Who can use it: BrowserSkill CLI users; the plan tool is also available in MCP and DSH integrations.

    How to use it: Run the bsk plan command with the latest CLI; enable the plan flow on the MCP or DSH side and send sequential actions.

    Tip: Validate the plan on its first run with short, reversible steps.

  • FeatureBrowserSkill

    BrowserSkill saves and replays verified flows with recipe

    The CLI adds the bsk recipe flow to save and rerun verified workflows. Supported reversible steps require confirmation, while ambiguous outcomes trigger a safety hold.

    Usage details

    Who can use it: BrowserSkill CLI; recipe tools can be enabled optionally in MCP and DSH.

    How to use it: Save a flow with bsk recipe and replay it by the same recipe name on the next run; enable recipe tools separately in MCP or DSH.

    Tip: If the outcome is ambiguous, check the status with bsk sync and ledger status before proceeding with new write operations.

  • CapabilityBrowserSkill

    BrowserSkill sends text-only changes as observation patches

    When the target policy is to change only visible text in place on a page, cooperate sync no longer resends the entire VOM. The CLI can merge the full observation as a patch onto a matching baseline and save it for offline review.

    Usage details

    Who can use it: When BrowserSkill 0.5.0 CLI, daemon, and extension are used together.

    How to use it: Run bsk sync --baseline <id> with the ID of the previous observation; review the merged result with bsk view and use --save-full when needed.

    Tip: Use only the exact ID of a previous observation as the baseline; take a fresh observation when there are structural changes.

  • CapabilityBrowserSkill

    bsk read adds bounded DOM reading for dynamic pages

    bsk read samples selected DOM fields, such as text, attributes, and form state, in a bounded way on dynamic pages. The same read operation is also available through the MCP and DSH surfaces.

    Usage details

    Who can use it: BrowserSkill CLI, MCP, and DSH users.

    How to use it: Run bsk read, specifying the required DOM fields; in scripts, declare dependencies explicitly for pre-action checks.

    Tip: Selecting narrow fields instead of reading the full page reduces both latency and context usage.

  • AvailabilityDSpine

    DSpine adds MLX inference support on Apple Silicon

    DSpine adds Apple Silicon inference support through mlx-lm and SGLang-MLX. The DSpine draft flow for Qwen3.8-27B is no longer limited to CUDA environments.

    Usage details

    Who can use it: Apple Silicon Macs; users with mlx-lm or SGLang-MLX installed.

    How to use it: Install the DSpine MLX package: pip install -e dspine_mlx; then run the mlx-lm or SGLang-MLX serving command in the MLX guide.

    Tip: On Mac, first measure memory usage with a small batch; do not carry over CUDA-specific SGLang options unchanged.

    ✓ Verified by official sourceModels: Qwen3.8-27B, Qwen3.8-27B-DSpine-b16Source: DSpine resmi READMESource: DSpine 12 Kasım resmi commit listesi
  • APIHy-Memory

    Coding/Working edits with embedding errors preserve old data

    When an embedding cannot be generated for new text in Coding or Working memory, the API now returns 503 memory_update_embedding_failed and leaves the old content and vector unchanged. A failed edit is no longer silently reported as successful.

    Usage details

    Who can use it: Hy-Memory 2.0.16 HTTP/API editing workflows.

    How to use it: Handle the 503 and memory_update_embedding_failed code in the edit response; fix the embedding provider before retrying the unsuccessful change.

    Tip: On the client side, treat 503 as an idempotent retry and inform the user that the old content was preserved.

  • SettingLoopForge

    Confirmation requested before handing a requirement to an SOP

    LoopForge has begun asking for user confirmation before handing a requirement to an SOP. This reduces unexpected automation when moving from chat to workflow execution.

    Usage details

    Who can use it: LoopForge chat and SOP workflows.

    How to use it: Route a requirement to an SOP; execution will not start until you accept the confirmation prompt.

    Tip: Keep confirmation enabled by default for workflows that touch production repositories.

  • RetiredTencent Cloud TokenHub

    Hy-Image-3.0 discontinued on TokenHub

    Tencent Cloud discontinued the Hy-Image-3.0 model with the hy-image-v3 API parameter on November 12, 2026, at 23:59:59 Beijing time; calls will be unavailable after removal. The official recommendation is to switch to Hy-Image-3.5-preview and change the model parameter in your code in advance.

    Usage details

    Who can use it: All Hy-Image-3.0 API calls using Tencent Cloud LLM Service TokenHub; the old parameter will be unavailable after removal.

    How to use it: Replace the hy-image-v3 model parameter in your code with hy-image-v3.5-preview, test the request according to the Hy Image Generation API guide, and check billing before deploying to production.

    Tip: Make the switch before the removal date and compare outputs from the old and new models using the same test images.

    ✓ Verified by official sourceModels: Hy-Image-3.0, Hy-Image-3.5-previewSource: Tencent Cloud Hy-Image-3.0 kaldırma duyurusuSource: Tencent Cloud model listesi
  • Plugin / IntegrationHy-Memory

    Optional PostgreSQL backend added for Hy-Memory business storage

    Hy-Memory 2.0.16 adds an optional PostgreSQL backend by moving Working/Coding experiences, facts, raw traces, and the Coding version chain behind MemoryRepository and ChainRepository. SQLite remains the default, and existing SQLite data is not migrated automatically.

    Usage details

    Who can use it: Hy-Memory 2.0.16 users; new installations that want PostgreSQL.

    How to use it: Install the hy-memory[postgres] extra package and configure the repository/backend as described in docs/business_storage.md; migrate existing SQLite data separately.

    Tip: Back up SQLite before migrating; the release does not perform automatic migration.

  • FeatureLoopForge

    Sandbox development servers open natively in the side browser

    LoopForge now supports displaying sandbox development servers in a separate, native side browser surface. This lets users inspect a local web app created while the agent is running within the same work experience.

    Usage details

    Who can use it: Users running sandbox development servers on the LoopForge platform.

    How to use it: Start a development server in a sandbox and open the side browser in the run view; the server interface will appear in the browser panel.

    Tip: Using the side browser instead of exposing the preview address keeps the local development workflow more secure.

  • Settingteamai-cli

    Teams can configure default resource thresholds with params.yaml

    teamai-cli added team-level configuration of thresholds and resource behavior through params.yaml in a default resource directory. Range and unknown-key validation rejects invalid parameters early.

    Usage details

    Who can use it: teamai teams and default resources that use params.yaml.

    How to use it: Add params.yaml to the relevant default directory, define supported threshold or hook parameters, and run the teamai refresh/pull flow.

    Tip: Test parameters with small changes and submit them to the team repository with comments.

  • CapabilityUniRL

    UniRL adds AROPD on-policy distillation for AR models

    UniRL added AROPD, a sampled-token on-policy distillation approach. The flow supports autoregressive model training with raw-gap advantage, gap_clamp, and actual token ratio metrics.

    Usage details

    Who can use it: UniRL AR training flows; configurations using the AROPD algorithm.

    How to use it: Update UniRL, select the AROPD algorithm in the AR training configuration, set sampling_temperature, and start a sample training run.

    Tip: Calibrate gap_clamp and sampling_temperature values in a small validation run.

  • SettingMemHub

    agent_max_tokens is now a persistent memory-features setting

    MemHub has added agent_max_tokens as a persistent memory-features setting for local models’ evolve and System 2 output. The value is kept within the 512–32000 range, and the legacy environment variable continues to work by default.

    Usage details

    Who can use it: Local model installations using MemHub 1.0.7.

    How to use it: Run memhub settings features set agent_max_tokens=6000 --yes; to reset it, use memhub settings features reset agent_max_tokens.

    Tip: Increase it when longer output is needed, but monitor local model memory use and latency.

    ✓ Verified by official sourceSource: MemHub 1.0.7 resmi sürüm commit’i
  • CapabilityAI-Infra-Guard

    Agent-Scan parses batch arrays in XML parameters

    Agent-Scan correctly parses batch array values passed within XML parameters. This prevents tool calls with multiple items from being omitted or misread during audits.

    Usage details

    Who can use it: AI-Infra-Guard Agent-Scan.

    How to use it: Rescan agent examples with XML parameters using the latest Agent-Scan; no additional conversion code is needed.

    Tip: Test batch-call examples in both single-item and array formats.

  • CapabilityLoopForge

    Tool and executor outputs are safely truncated at rune boundaries

    LoopForge runtime truncates tool messages and executor output at rune boundaries without splitting UTF-8 characters. This reduces the chance of garbled characters in long output displays.

    Usage details

    Who can use it: LoopForge runtime and chat runner.

    How to use it: No additional settings are required; generate a long tool output and use the truncated view in the chat history.

    Tip: For debugging that requires the full output, inspect the run artifact instead of the truncated chat message.

  • CapabilityHy-Memory

    Chroma content updates preserve the existing vector

    When only the payload content is updated in Chroma, the stored vector is no longer regenerated using the default model. In addition, the tag_ners field is preserved as a list through read-write cycles.

    Usage details

    Who can use it: Hy-Memory 2.0.16 and applications using Chroma storage.

    How to use it: Use update_payload to change content; the existing embedding is preserved when only the payload changes.

    Tip: If you intend to change the embedding, run an explicit re-embed workflow rather than silently relying on a content update.

  • APIHy-Memory

    Coding inference reports OUTPUT_TRUNCATED when cut off at the token limit

    Hy-Memory 2.0.16 explicitly reports Coding inference cut off by the token limit with the OUTPUT_TRUNCATED result. The client no longer has to treat an incomplete inference as a complete result.

    Usage details

    Who can use it: Hy-Memory 2.0.16 Coding inference API.

    How to use it: Check the response status and, when you see OUTPUT_TRUNCATED, rerun with a higher output token budget or a shorter prompt.

    Tip: Monitor the truncation rate; frequent truncation may require adjusting the provider’s output token parameter.

  • CapabilityDSpine

    DSpine generalizes batch, sampling, and dynamic verification settings

    The DSpine serving layer has generalized batch and per-request sampling behavior; the dynamic verification tuning workflow can now also be applied to different workloads. These options can be used in the same workflow for Qwen3.8-27B draft serving.

    Usage details

    Who can use it: DSpine serving and SGLang-based Qwen3.8-27B inference.

    How to use it: Update DSpine, generate a table suited to your workload using the SPS tuning guide, and add --speculative-dspark-sps-table-path sps.json to the serving command.

    Tip: Generate the SPS table using your actual batch distribution; don’t blindly reuse the same table at different concurrency levels.

    ✓ Verified by official sourceModels: Qwen3.8-27B, Qwen3.8-27B-DSpine-b16Source: DSpine 12 Kasım resmi commit listesi
  • APIBrowserSkill

    get-html JSON output writes HTML to a file and returns its path and size

    bsk get-html --json --out now writes HTML content to the specified file instead of printing it to standard output, and returns the file path and size in JSON. This makes it easier to pass large HTML responses through a command pipeline.

    Usage details

    Who can use it: BrowserSkill CLI 0.5.0.

    How to use it: Run bsk get-html --json --out page.html; pass the path and size fields in the JSON response to the next step.

    Tip: For large pages, use --out instead of stdout to avoid context and pipeline overflow.

  • CapabilityHy-Memory

    A faulty pooled client is not retired until in-flight requests finish

    The Hy-Memory LLM client pool does not retire a client that has encountered an error until its ongoing requests have completed. This prevents a single error from unnecessarily interrupting concurrent writes.

    Usage details

    Who can use it: Hy-Memory 2.0.16 LLM client pool and concurrent write workflows.

    How to use it: No additional settings are required; upgrade to 2.0.16 and run the concurrent memory write workflow as usual.

    Tip: When monitoring the provider pool, track individual errors and completed concurrent requests in separate metrics.

  • CapabilityAI-Infra-Guard

    Hex-suffixed tool-call tags are normalized

    MCP, skill, and agent scanners convert tool-call tags with hex suffixes into a common format. This prevents provider-specific variants from disrupting scan parsing.

    Usage details

    Who can use it: AI-Infra-Guard mcp-scan, skill-scan, and agent-scan.

    How to use it: No additional settings are required; install the latest scanner version and use the same scan commands.

    Tip: Compare scan logs before changing raw tool-call tags in provider templates.

  • SettingHy-Memory

    Hy-Memory logs can be routed to stderr

    Hy-Memory 2.0.16 adds the MEMORY_LOG_STREAM=stderr option to separate INFO and DEBUG logs from stdout in stdio hosts. The default behavior is unchanged.

    Usage details

    Who can use it: Stdio integrations using Hy-Memory 2.0.16.

    How to use it: Set the MEMORY_LOG_STREAM=stderr environment variable before running; keep stdout clear for protocol data.

    Tip: Prefer this setting in hosts that use JSON-RPC or another stdout protocol.

  • AvailabilityMemHub

    Linux adds a Secret Service keyring backend and hang-resistant D-Bus access

    A Linux Secret Service backend was added to MemHub 1.0.7’s frozen Core package. Running D-Bus keyring read, write, and delete operations in a terminable subprocess prevents sdk-migrate and model key configuration from hanging.

    Usage details

    Who can use it: Linux x64/arm64 MemHub Runtime.

    How to use it: Keep the Secret Service provider running on Linux and use the memhub models configure or sdk-migrate command; the current package selects the backend automatically.

    Tip: On headless servers, first verify Secret Service access with a small keyring test.

    ✓ Verified by official sourceSource: MemHub 1.0.7 resmi sürüm commit’i
  • FeatureMemHub

    MemHub reports stranded write status with the actual count

    In MemHub’s memory-processing view, stranded writes with a missing or expired lease are no longer shown as zero. The web client also displays a null status as unknown, preventing false healthy-status reports.

    Usage details

    Who can use it: MemHub 1.0.7 memory-processing and web status screen.

    How to use it: Open memory-processing or the relevant dashboard status, and monitor stranded_depths and unknown-status rows.

    Tip: If the stranded count rises, inspect worker leases alongside the time of the last successful Core write.

    ✓ Verified by official sourceSource: MemHub 1.0.7 resmi sürüm commit’i
  • Plugin / IntegrationMemHub

    OpenCode V2 history and MCP configuration imported

    MemHub 1.0.7 imports session history from OpenCode 2’s new store structure; the capture plugin and MCP configuration also use V2’s mcp.servers format. History previously imported under V1 is not imported again.

    Usage details

    Who can use it: Users of MemHub 1.0.7 and OpenCode 2.

    How to use it: Connect the OpenCode V2 store, update the capture plugin, and define MCP settings in mcp.servers format; MemHub will run the history import.

    Tip: To prevent duplicate records when migrating from V1 to V2, do not delete the old import markers.

    ✓ Verified by official sourceSource: MemHub 1.0.7 resmi sürüm commit’i
  • FeatureLoopForge

    The missing-key notice clears when a Quick Access key is saved

    LoopForge chatbot clears the old missing-key notice when a missing key is saved through Quick Access. Users no longer keep seeing the same warning by mistake after adding the key.

    Usage details

    Who can use it: Workspaces using LoopForge chatbot and Quick Access.

    How to use it: Open Quick Access in the chatbot, save the missing provider key, and refresh the chat panel; the notice clears automatically.

    Tip: After adding a key, check that the provider status has cleared before starting a new chat.

  • FeatureLoopForge

    You can install a missing CLI from a chat error

    When a chat run fails because a CLI is missing, LoopForge offers an option to install the missing tool from the error flow. Users can install the missing dependency without having to search through settings.

    Usage details

    Who can use it: LoopForge chat runner and supported built-in CLIs.

    How to use it: When a CLI-missing error appears, select the installation action on the conversation error card; rerun the same request after installation.

    Tip: After installation, also verify the provider key and CLI version.

  • Plugin / IntegrationLoopForge

    Per-host user keys now supported for SSH repositories

    LoopForge Git integration now supports using different user keys per host for SSH repositories. Teams connecting to multiple Git servers are no longer tied to a single shared key.

    Usage details

    Who can use it: LoopForge users connecting Git repositories over SSH.

    How to use it: Connect SSH repositories and define the user key for the relevant host in your SSH configuration; LoopForge uses the host match when cloning.

    Tip: Clearly separate host aliases in ~/.ssh/config to prevent the wrong account from being used.

  • FeatureLoopForge

    Completed execution history can be force-deleted

    LoopForge now offers an option to force-delete settled execution history records after confirmation. This gives administrators control over completed records that get stuck during cleanup.

    Usage details

    Who can use it: Users managing LoopForge execution history.

    How to use it: Open the menu for a settled execution history record, select the force delete option, and confirm.

    Tip: Before deleting, export any logs needed for debugging or auditing.

  • APIAI-Infra-Guard

    Scanners pass the optional reasoning effort value to the model

    AI-Infra-Guard passes the optional reasoning effort value to the model API in mcp-scan, skill-scan, and agent-scan workflows. This makes it possible to adjust the reasoning budget per scan with supported providers.

    Usage details

    Who can use it: AI-Infra-Guard scanners and model providers that accept reasoning_effort.

    How to use it: Set a supported REASONING_EFFORT value in the scan environment and rerun the mcp-scan, skill-scan, or agent-scan command.

    Tip: Limit costs by using a higher reasoning level only for complex tool or skill checks.

  • CapabilityHy-Memory

    Repeated user prompt gets a new chain version

    In the Coding workflow, a user prompt repeated in a later turn now gets its own chain version record instead of being merged with the previous turn. Version history and turn separation are preserved.

    Usage details

    Who can use it: Hy-Memory 2.0.16 Coding experience and chain history.

    How to use it: Resubmit the same user prompt in a later Coding turn; the current SDK generates a new chain version.

    Tip: When comparing versions, store the turn and chain version ID alongside the prompt text.

  • CapabilityAI-Infra-Guard

    Local asynchronous tool execution errors no longer crash scans

    Scanners now handle errors from asynchronous execution of local tools in a controlled manner. An error from a single local tool no longer causes the entire scan process to crash unexpectedly.

    Usage details

    Who can use it: AI-Infra-Guard scanner workflows and local tool execution.

    How to use it: Start a scan as usual with the latest version; asynchronous local tool errors will be reported in the scan results.

    Tip: Instead of silently ignoring error reports, save them along with the tool name and reproduction steps.

  • CapabilityLoopForge

    Length limit removed for scheduled task prompts

    The previous fixed limit on scheduled task prompt length in the LoopForge scheduling workflow has been removed. More extensive context and instructions can now be passed to scheduled runs.

    Usage details

    Who can use it: Workspaces using LoopForge scheduled tasks.

    How to use it: When creating a scheduled task, enter the long prompt in the regular prompt field; the old length cutoff no longer applies.

    Tip: Even with the limit removed, divide the prompt into sections to keep task costs and execution time in check.

  • CapabilityAI-Infra-Guard

    Agent-Scan preserves non-object HTTP error responses

    Agent-Scan now preserves HTTP error bodies that are not in object form. Provider error diagnostics are no longer limited to endpoints that return JSON objects.

    Usage details

    Who can use it: AI-Infra-Guard Agent-Scan and HTTP-based providers.

    How to use it: Scan a provider that returns HTTP error responses with the latest Agent-Scan; the error body will be preserved in the result record.

    Tip: Add the raw error body to the diagnostic record after ensuring it contains no sensitive information.

  • CapabilityAI-Infra-Guard

    Agent-Scan preserves prompt escapes in provider templates

    Agent-Scan preserves prompt escape characters in provider templates, passing the correct text to the model and the scan. This reduces false findings caused by altered prompts in templates containing escapes.

    Usage details

    Who can use it: AI-Infra-Guard Agent-Scan provider templates.

    How to use it: Use the templates with the latest version; rescan a provider template containing custom escape characters.

    Tip: Compare template changes against both the raw file and the resolved prompt output.

  • FeatureLoopForge

    AI orchestrator build animations are back

    AI orchestrator build animations are visible again in the LoopForge web interface. Users can visually track that the process is ongoing while a workflow is being created.

    Usage details

    Who can use it: LoopForge web interface.

    How to use it: Create a workflow with AI orchestrator; watch the animation status in the interface while it builds.

    Tip: For long builds, wait for the animation to finish without refreshing the page.

  • FeatureLoopForge

    Activity animation shown on running SOP nodes

    The LoopForge web interface has begun showing activity on running SOP nodes with animation. Users can more easily track which node is processing.

    Usage details

    Who can use it: LoopForge web SOP view.

    How to use it: Run an SOP and watch the animated status indicators on running nodes in the workflow graph.

    Tip: For long workflows, use the animation alongside node logs to distinguish waiting from errors.

  • CapabilityHy-Memory

    Memory Strength supports different timestamp formats

    Memory Strength idle decay now correctly converts timezone-aware dates, ISO strings with offsets, and epoch integers. Unreadable timestamps continue to use a safe default.

    Usage details

    Who can use it: Flows using Hy-Memory 2.0.16 Memory Strength and bump_access.

    How to use it: Run the compute_strength or bump_access flow in the updated SDK with existing memory timestamp values; no additional conversion is needed.

    Tip: For new integrations, avoid ambiguous local times by using UTC-based epochs or explicit offsets.

  • FeatureLoopForge

    The web interface shows the commit of the platform image in use

    The LoopForge web interface now displays the commit information for the running platform image in a corner. This makes version troubleshooting easier in environments using images with the same tag.

    Usage details

    Who can use it: LoopForge web platform.

    How to use it: Open the platform web interface and read the image commit indicator in the corner; share this value in support requests.

    Tip: When reporting a behavior change, record the image commit as well as the version tag.

  • Availabilityteamai-cli

    Team scripts run with hidden console windows on Windows

    teamai-cli now runs Windows team scripts without opening a visible console window. This reduces desktop interruptions for background automation.

    Usage details

    Who can use it: teamai team scripts on Windows.

    How to use it: Run the teamai team script as usual on Windows; the current version hides the console window automatically.

    Tip: Enable script logs separately when debugging; a hidden window can make error output invisible.

  • Capabilityteamai-cli

    Repeated pull operations make fewer file-system calls

    teamai-cli reduced file-system calls in repeated pull operations by about 59% and can deliver a copy of a tool installed by another tool. This reduces pull time and I/O load in installations that use the same resources.

    Usage details

    Who can use it: teamai-cli pull and multi-tool installations.

    How to use it: Run teamai pull with the current version; if the same copy was installed previously, the tool installation reuses it.

    Tip: In CI environments, compare repeated pull measurements separately with and without cache clearing.

  • Capabilityteamai-cli

    Host-blocked calls are no longer counted as tool errors

    In the teamai-cli contribute flow, a call blocked by the host is no longer reported as a failed tool execution. This distinction separates actual tool errors from environment policy blocks.

    Usage details

    Who can use it: teamai-cli contribute and host policy restrictions.

    How to use it: Run the Contribute flow with the current version; when a host block occurs, treat the result as a policy block.

    Tip: Check the host permission policy before retrying blocked calls.

Watchlist

Kimi K3.1 Pro

46
Moonshot AI · Kimi

Dropped out of the TOP 10 and is now No. 11 after Gemini 4 Hydrogen entered; it remains the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. On November 12, Kimi Code gained Tower `artifact` tasks and experimental prompt suggestions.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; ranked No. 12 with 44 points. In preview on Google Agent Platform as `muse-wonder-1.0`.

Claude Haiku 5.5

36
Anthropic · Claude

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

Muse Ember

36
Meta · Meta AI

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

Grok 4.7

35
xAI · Grok

The current model in the Grok app and API. Grok 4.8 is expected to become available to some Grok Bot users; it has no official API ID, price, or Artificial Analysis score yet. Grok Bot 1.0 was released on November 11.

Tencent Hy4

35
Tencent · Hy

Available through the API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, free on WorkBuddy and CodeBuddy through November 15. The old Hy-Image-3.0 image model was removed from TokenHub on November 12.

Sources