Anthropic · Claude
No notable change since the last edition.
No changes in the TOP 10 today either. The rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are the same as yesterday: Claude Fable 5.2 leads with 65 points, followed by Gemini 4 Carbon at 60, Claude Opus 5.5 at 59, and GPT-6 Astra at 57. GPT-6.1 Sol and Claude Sonnet 5.5 are tied at 55, Gemini 4 Hydrogen has 54, GPT-6.1 Terra has 51, and Gemini 4 Argon and Muse Wonder Full have 50 points. There is one change on the watchlist: Alibaba’s Qwen4 Max, released today, entered 13th place with 43 points. This score appears on Artificial Analysis’s model page and may not yet be reflected in the main table. Tencent Hy4 (35) has therefore left the six-model watchlist; Grok 4.7 ranks ahead on Artificial Analysis. Kimi K3.1 Pro is in 11th place with 46 points and remains the strongest open-weight model. LMArena’s live page was again blocked by a CAPTCHA; the latest accessible snapshot is dated November 11.
The day’s main story is Qwen4-Max. Alibaba’s new flagship is built on the new Qwen4-series architecture: a 2.4-trillion-parameter backbone, 71 billion active parameters per token, and a separate 328-billion-parameter N-gram embedding layer. The model accepts text, image, and video input and offers a 1-million-token context window. In Alibaba’s own Work360 test, the score rises from 47.74 to 52.69, while the average cost per task falls from $4.43 to $1.67. The model, which opened on November 11 with only a Token Plan preview, is now available in six regions as a pay-as-you-go API in Model Studio under the `qwen4-max` identifier. It is compatible with OpenAI Chat Completions, Responses, Anthropic Messages, and DashScope. In Singapore, pricing per 1M tokens is $1.50 for input and $4.50 for output, below Qwen3.8-Max’s $2 / $6 pricing. Cached reads cost $0.075. There is also a free quota of 1 million tokens, valid for 90 days from Model Studio activation or the model’s release. The reasoning level is set to low, high, or xhigh via `reasoning_effort`; the default is high. Alibaba also introduced Live Delivery, an approach that delivers an initial usable result early on long tasks and improves it in subsequent versions, and open-sourced Qwen-Live-Delivery-Plugin.
Alibaba’s other entries are from Qwen Code. In v0.27.4, released with CLI, Desktop, and TypeScript SDK 0.1.34, Managed Agent sessions can temporarily access local folders you approve in the browser; folder paths and credentials are not stored. Interactive Web Terminal, accessed in the same sessions with a short-lived ticket, has also arrived. Feishu custom bots and WeCom group webhooks have been added as new send-only channels; long responses are delivered in sequential chunks instead of being truncated. The Goals page brings multiple trusted workspaces together in a single view. `/memory` controls have returned to the CLI. Typed APIs for importing Codex memory have been added to the SDK. Hook executions are now recorded in session history on a per-turn basis. Canceled runs emit the `loop.cancelled` event. If the fast classifier times out in Auto Mode, the session model evaluates the action once more. The November 15 nightly release is testing the ability to run Managed Agent in another workspace, adding DingTalk to the Managed Channel path, and adding a workspace filter to the sidebar.
Tencent’s 13 entries are from developer tools. In teamai-cli, workflow steps can now be delegated to connected drivers defined in `drivers/<name>.yaml` and to external runtimes (`teamai workflow start <workflow> --driver <name>`). Concurrent MCP manifest writes are queued using a file lock. `teamai doctor --json` reports a malformed `roles.yaml` file as an error and keeps JSON output clean. LoopForge’s `bind-cli` tool no longer overwrites an unreadable crontab or terminates the wrong process because of a stale PID file. AI-Infra-Guard fixes include case-insensitive recognition of the end tool in Agent-Scan and MCP-Scan, a CIDR scan loop on the final IPv4 address, and Host header errors. Meta made v130 software mandatory for its Adventurer, Fury, Starfire, Nova, and Capri-style AI glasses. This version fixes overheating after exposure to moisture or sweat; after November 16, some features will be blocked on glasses without v130 installed. At Moonshot, Kimi Code Web Canary has started drawing Mermaid swimlane diagrams. There were no new user-facing changes on the official channels of OpenAI, Anthropic, Google, xAI, DeepSeek, Z.ai, and Xiaomi between November 14 and 16. Codex’s 0.171.0 alpha tags have no release notes. The domain migration for Gemini Notebook, which takes effect today, was already covered in the November 6 entry. A total of 42 new developments were added to the archive today.
The ten best unique models by intelligence index; each model is counted at its highest-scoring setting.
Released on November 9 and leads with 65 points. Available on all paid Claude plans; `claude-fable-5-2` in the API, $8 / $40, with thinking always set to adaptive. On November 12, the Claude Meetings beta (Google Meet first) opened to Team and Enterprise plans.
`gemini-4-carbon` in Gemini API and AI Studio (paid tier only; $2 / $10 until 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 on Ultra.
Anthropic’s flagship Opus model; available on Claude Pro/Max/Team/Enterprise and in the API. Claude Code 2.1.322 added fast mode support on Vertex (global, US, EU).
OpenAI’s most powerful GPT-6 model. In Agents SDK Python v0.24.0-alpha.7, the default reasoning effort for this model went from `low` to `medium`.
The default model on paid ChatGPT plans; API pricing is $2 / $10. The default reasoning effort became `medium` in an Agents SDK alpha release.
The most powerful model also available on claude.ai’s free plan; tied with GPT-6.1 Sol at 55 points.
Released on November 12; scores 54 with the xhigh setting, at a cost of $3.16 per task ($5.48 for Carbon). `gemini-4-hydrogen` is also available on the free API tier, with paid pricing of $1.50 / $7.50. Available through Firebase AI Logic and Android Studio. According to Google’s plans page, Hydrogen is also the top model in the Gemini app for Google AI Plus subscribers.
Released on October 26 in ChatGPT Work, Codex, and the API; focused on image and computer use. The administrator preview of Dots agents continues until November 18.
Available on Google AI Pro/Ultra and the paid API; the $2 / $10 introductory price is available until November 30. In Antigravity, Orchestrator (Argon) is available to both Pro and Ultra subscribers.
Meta’s most powerful Muse model; no public API. This week, the Muse app added a feature for paid subscribers in the US to call businesses, as well as the Muse from Meta Chrome extension.
Choose a region in the Model Studio console and create an API key. If you’re using it from Türkiye, Singapore, where the free quota is available, is a practical place to start. In your existing OpenAI Chat Completions, Responses, or Anthropic Messages code, change only the endpoint to the Model Studio address for that region and the model name to `qwen4-max`. Use `reasoning_effort: "low"` for simple tasks and `"xhigh"` for difficult ones; the default is `high`. Put parts repeated in every request, such as long system prompts and tool definitions, in the explicit cache (5 minutes or 1 hour); cached reads cost $0.075 per 1M tokens. If you use Qwen3.8-Max, run the same test set on both models and compare cost and quality; list pricing drops from $2 / $6 to $1.50 / $4.50.
Alibaba · QwenUpdate with `npm install -g @qwen-code/qwen-code@latest` and verify version 0.27.4 with `qwen --version`. When Managed Agent asks for access to local folders, choose only the project folder for that task, not your home directory; permission is temporary, but everything in the folder is readable for the duration of the task. Open Web Terminal only in your own session and close it when you’re done. To send results to your team, choose Feishu Webhook or WeCom Webhook in Channel connections. Treat the webhook URL like a secret key; don’t put it in a repository or shared files, and enable the signing secret in Feishu as well. You can manage memory settings at the user or project level with `/memory` in the CLI.
Alibaba · QwenIf you have Meta AI glasses in the Adventurer, Fury, Starfire, Nova, or Capri style, open the Meta AI app on your phone. Select the glasses from the Glasses or Devices tab, go to Device settings, and check the software version. If it’s below v130 and automatic updates are off, start the update manually; do this while the glasses are charging and near your phone. According to Meta, some features will be blocked on glasses without v130 installed after November 16. The update also fixes overheating after exposure to moisture or sweat; don’t put it off if you use them during exercise.
Meta · Meta AIWhat changed in this edition for every AI in the TOP 10 and on the watchlist.
No notable change since the last edition.
No notable change since the last edition.
No notable change since the last edition.
New in this edition · 1
Meta says the v130 software resolves an overheating issue after exposure to excessive humidity or sweat for the Adventurer, Fury, Starfire, Nova, and Capri styles. If v130 or later is not installed after November 16, 2026, certain features will be blocked.
Who can use it: Meta Glasses in the Adventurer, Fury, Starfire, Nova, and Capri styles; v130 or later required after 2026-11-16.
How to use it: In the Meta AI mobile app, go to the Glasses or Devices tab, select your glasses, open Device settings, and install v130 or later if automatic updates are turned off.
Tip: Check the version number in the app before traveling or heavy use, and keep automatic updates enabled.
New in this edition · 1
In the official web deployment commit dated November 16 in the MoonshotAI/kimi-code repository, a `swimlanesDiagram` parser asset not present in the previous bundle was added. This shows that Mermaid `swimlane` diagrams can be processed in the daily Kimi Code Web Canary build; the PR description says this is only a Canary web UI sync and does not include a separate changeset.
Who can use it: Daily Kimi Code Web Canary build; availability in the stable release has not been announced separately.
How to use it: Add a Mermaid code block to a message in Kimi Code Web Canary and submit a `swimlane` diagram; check the rendered diagram in the response.
Tip: To keep complex processes readable, assign each lane to a single team, role, or stage, and keep the number of lanes limited.
New in this edition · 27
The Model Studio pricing table lists qwen4-max at $1.50 per 1 million input tokens and $4.50 per 1 million output tokens in Singapore. The list price for other international deployments is $1.413 for input and $4.24 for output; cached reads also have a discounted unit price.
Who can use it: Model Studio's regional, pay-as-you-go API pricing; a free quota of 1 million tokens is also listed for Singapore.
How to use it: Check the deployment you use and the qwen4-max row on the Model Studio pricing page; route the call to the endpoint in the same region.
Tip: Reduce input costs by using explicit or implicit caching for repeated system prompts and tool definitions.
Qwen4-Max supports the Live Delivery approach, which delivers an initial usable output early for long-running tasks and improves it in subsequent versions. Users can provide guidance throughout the process; Qwen-Live-Delivery-Plugin has also been open-sourced.
Who can use it: Long-running Agent scenarios with Qwen4-Max; the plugin is available via GitHub.
How to use it: Install the plugin using the Qwen-Live-Delivery-Plugin link in the Qwen announcement, then run the long task so it accepts interim deliveries and user approvals.
Tip: Use the first interim delivery as a small version to test acceptance criteria; correcting the direction early makes subsequent work less costly.
Alibaba unveiled Qwen4-Max as its flagship, built on the new Qwen4 architecture with 2.4 trillion parameters. The model targets native vision, a 1 million-token context window, professional tool use, and long-running task execution.
Who can use it: Unveiled in the Qwen announcement; available in Model Studio under the qwen4-max model ID.
How to use it: Set the model parameter to qwen4-max in a Model Studio call and send the request using one of the supported API examples.
Tip: For long tasks, control costs and latency by selecting relevant documents and tool outputs rather than filling the 1 million-token context window.
qwen4-max was added to Model Studio's new model list, separate from the qwen4-max-preview entry that was Token Plan-only on November 11. The model page lists compatibility with OpenAI Chat Completions, Responses, Anthropic Messages, and DashScope across deployments in China, Singapore, Frankfurt, Virginia, Tokyo, and Hong Kong.
Who can use it: Model Studio regions in mainland China and internationally: Singapore, Frankfurt, Germany; Virginia, US; Tokyo, Japan; and Hong Kong.
How to use it: Choose an eligible regional endpoint in Model Studio, enter qwen4-max in the model field, and use one of the Chat Completions, Responses, Anthropic Messages, or DashScope protocols.
Tip: Choose the deployment closest to where user data is located; check before making the call because changing regions can affect latency, data residency, and price.
Managed Agent sessions can temporarily access local folders granted by the browser user. Credentials and folder paths are not stored permanently.
Who can use it: Qwen Code v0.27.4 Managed Agent sessions; access is limited to local directories granted by the user.
How to use it: When a local file is needed in a Managed Agent session, approve the folder access request; authorize only the necessary directory temporarily.
Tip: Reduce the visibility of sensitive files by granting access only to the subfolder required for the task instead of the repository root.
Managed Agent can open an interactive Managed Web Terminal authenticated with a short-lived identity ticket. The runtime remains active while connected to the terminal.
Who can use it: Qwen Code v0.27.4 Managed Agent sessions and Web Shell users.
How to use it: Open a Managed Agent session and select the Managed Web Terminal entry in the session; use the terminal through the temporary connection that is created.
Tip: For long-running commands, check the session’s workspace and permission scope before leaving the terminal connection open.
Qwen Code CLI v0.27.4 has been released, and the release notes do not mention any known backward incompatibilities. The release includes new user-facing features for Managed Agent, channel delivery, Goals, memory, permissions, and the execution loop.
Who can use it: Qwen Code CLI users; the release is available on the QwenLM/qwen-code GitHub release page.
How to use it: Upgrade your installation to Qwen Code CLI v0.27.4 and verify the version in your current workspace.
Tip: Before upgrading in a production workflow, verify channel connections, Managed Agent settings, and experimental loop flags in a short test session.
When the fast classifier times out and the identifiers do not match, the session model can review the action once. The decision is to allow or block the action according to the current permission policy.
Who can use it: Qwen Code v0.27.4 Auto Mode; workflows where the current permission policy and session model are available.
How to use it: Enable Auto Mode as usual; when the fast classifier times out, approve or reject the second assessment through the existing permission flow.
Tip: For destructive operations, use this fallback with explicit user approval rather than automatic permission.
The CLI brings back local /memory controls and User/Project file or managed-folder actions. Workspace settings persist across restarts.
Who can use it: Qwen Code CLI v0.27.4 users; local or managed workspaces.
How to use it: Open the /memory command in the CLI; select a memory action under the User, Project, or managed-folder scope, and save the workspace setting.
Tip: Put notes that should not be shared across the team under User scope instead of Project scope.
The Workspace-bound SDK provides typed APIs to preview, submit, review, cancel, and retry Codex memory imports. This allows the import flow to run without being tied to the user interface.
Who can use it: Embedded Qwen Code applications using the Workspace-bound TypeScript SDK.
How to use it: Call the workspace-bound memory import API in the SDK; preview first, then submit, and review, cancel, or retry based on the status result.
Tip: Save the preview result before submission to keep a verifiable record of which entries were imported.
The nightly release adds the DingTalk adapter to the Managed Channel execution path. This expands availability in the Managed Agent channel path, separately from the previously recorded DingTalk webhook channel.
Who can use it: Qwen Code v0.27.3-nightly.20261115 Managed Channel nightly releases.
How to use it: In the nightly release, select the DingTalk adapter in the Managed Channel configuration and authorize the channel using your organization’s DingTalk connection details.
Tip: Before replacing an existing webhook channel, verify the new Managed Channel path in a single test group.
Qwen Code adds a send-only connection method for Feishu custom bot webhooks. The secret URL, optional signing secret, channel selection, and scheduled group delivery can be configured.
Who can use it: Managed Channel setups using a Feishu custom bot webhook.
How to use it: In the Channel connection methods, select Feishu Webhook, enter the bot’s secret URL and, if needed, the signing secret; then enable the channel and scheduled delivery.
Tip: Manage the signing secret separately from the webhook URL as a secret value, and verify it with a short notification in a test channel.
The Web Shell Goals page combines goals from multiple trusted workspaces in a single view and displays workspace badges. You can select a target workspace when creating a new goal.
Who can use it: Qwen Code Web Shell users with multiple trusted workspaces.
How to use it: Open the Goals page in Web Shell to view trusted workspaces, then select a target workspace when creating a new goal.
Tip: Check the workspace badges when creating a goal to avoid confusing goals with the same name.
Qwen Code added Hook execution records for each turn, summaries, and read-only Hook pages. SDK access and signed history snapshots are also provided.
Who can use it: Qwen Code v0.27.4 session history and SDK integrations.
How to use it: Open the relevant turn in session history and view the Hook record; in an application using the SDK, query the signed history snapshot.
Tip: For production audits, store the signed snapshot alongside the Hook summary for later comparison.
Cancelled work emits the loop.cancelled event and is terminated within the 10-second limit at most. Completed tool-group results are preserved when possible.
Who can use it: Qwen Code v0.27.4 Agent Loop and event-consuming integrations.
How to use it: Update your event stream to handle the loop.cancelled event; check for batch results preserved after cancellation.
Tip: In workflow automation, use the cancellation event to safely save previous results, not as a trigger to retry.
The nightly channel flow collects and delivers a summary of linked PR comments that accumulated during an outage when service resumes. This makes the backlog visible instead of silently skipping comments after a long outage.
Who can use it: Qwen Code v0.27.3-nightly.20261115 linked PR comment channels.
How to use it: Keep the linked PR channel enabled on the nightly release; after reconnecting, check the channel’s backlog summary and delivery of new comments.
Tip: Review the backlog summary before replying automatically after an outage so you don’t confuse an old comment with a new request.
The November 15 nightly release enables Managed Agent to run in a workspace outside the session’s default workspace. This is an execution-scope option separate from local file access and Web Terminal changes.
Who can use it: Nightly Managed Agent workflows in Qwen Code v0.27.3-nightly.20261115; it is not separately listed in the stable v0.27.4 notes.
How to use it: When creating a Managed Agent task in the nightly release, select the external workspace as the execution context and approve access.
Tip: Limit external workspaces to trusted directories; verify the session’s expected repository and file scope before the first tool call.
Qwen Code Desktop v0.27.4 brings the same user-facing features as CLI v0.27.4 to the desktop package. The desktop release also includes temporary recording of client file systems in Managed Agent sessions and new channel features such as Feishu.
Who can use it: Qwen Code Desktop users on macOS, Windows, and Linux; packages are available in the QwenLM/qwen-code GitHub release.
How to use it: Install the desktop-v0.27.4 package for your operating system, or update your existing desktop installation to this version.
Tip: After the desktop update, review local folder permissions and channel delivery targets for Managed Agent.
Qwen Code TypeScript SDK v0.1.34 was released and bundles Qwen Code CLI v0.27.4 as its source release. This is a new SDK package following v0.1.33, recorded on November 13.
Who can use it: Embedded applications and integrations consuming the TypeScript SDK.
How to use it: Upgrade the SDK dependency to v0.1.34; then verify in the build output that the bundled CLI version is 0.27.4.
Tip: Pair the SDK upgrade with build tests for typed API calls that touch memory and channel delivery.
When Managed Channel responses exceed the limit, they are split into ordered delivery segments instead of being silently truncated. If the segment limit is also exceeded, users see a truncation notice.
Who can use it: Qwen Code v0.27.4 Managed Channel deliveries, including Feishu and WeCom webhook channels.
How to use it: Configure the channel as usual; segmented delivery is applied automatically to long responses without additional settings.
Tip: For very long reports, still prefer delivering a summary and attachment; this reduces the chance of hitting the segment limit.
In the nightly release, the verify-pr flow expects a boundary declaration and takes ownership of PR watch status. This clearly separates verification scope from monitoring responsibility.
Who can use it: Qwen Code v0.27.3-nightly.20261115 verify-pr flows.
How to use it: Provide the requested boundary declaration when running verify-pr; stop using legacy channel-based watch settings, since the verify-pr flow now manages PR watch status.
Tip: Clearly specify the branch, PR, and base revision to be reviewed in the boundary declaration; this reduces ambiguity about monitoring scope.
Qwen Code has added a send-only channel that can send notification outputs, prompt results, and supported scheduled task results to a WeCom group. The connection method offers a group URL, channel controls, and scheduled delivery options.
Who can use it: Qwen Code Managed Channel setups with a WeCom group webhook configured.
How to use it: Select WeCom Webhook from the Channel connection methods, enter the group webhook URL, and choose a notification or scheduled delivery target.
Tip: To keep long task results from filling up the group feed, select only completion and error notifications for scheduled delivery.
Web Shell Goal status messages now show a one-line summary of the goal by default, with details available through the Expand details action. This keeps long goals from filling the chat stream.
Who can use it: Qwen Code v0.27.4 Web Shell Goal users.
How to use it: While a Goal is running, select Expand details in the status line to view the full goal.
Tip: For long goals, use a concise, distinctive opening sentence; the summary will remain clearer in compact views.
The Nightly release makes the resolved Agent Loop path visible on every Qwen Code surface. Users and developers can see which execution path is being used without inspecting experimental shared loop settings.
Who can use it: Qwen Code v0.27.3-nightly.20261115 surfaces: CLI, ACP, Managed, and Web Shell workflows.
How to use it: Start a Nightly session and check the resolved Agent Loop path in the execution status or diagnostics view.
Tip: When reporting behavioral differences, record the visible Agent Loop path as well as the model name.
The Qwen Code nightly release adds workspace filtering to the Web Shell sidebar. Users with many workspaces can show sessions from only the selected workspaces.
Who can use it: Qwen Code v0.27.3-nightly.20261115 Web Shell nightly users.
How to use it: Open the workspace filter in the Web Shell sidebar and select the workspace you want to view.
Tip: Filter for frequently used workspaces to reduce the risk of running an operation on the wrong project when searching for a session.
The Web Shell nightly release ensures voice dictation is finalized before a message is sent. This prevents the last words from being lost in the race between the end of dictation and sending.
Who can use it: Qwen Code v0.27.3-nightly.20261115 Web Shell voice input flow.
How to use it: Start voice dictation in Web Shell, finish dictating, and send the message; check that the final transcription has been entered in the input field before sending.
Tip: Especially for short commands, click the send button after the final text appears, not immediately after dictation stops.
When native UUID support is unavailable, Qwen Code now generates identities to keep terminal tabs, web previews, local commands, and slash commands working over plain HTTP.
Who can use it: Plain HTTP Web Shell or development environments without native UUID support.
How to use it: Run Qwen Code in a plain HTTP environment; if native UUID is unavailable, the version automatically generates the required identities.
Tip: Still use TLS in production for authentication and network security; treat this change only as a compatibility fallback.
No notable change since the last edition.
In 11th place; still the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. The stable release of Kimi Code 2.6.2 fixed slow startup in long sessions and stalls with empty responses.
Meta’s standard Muse Wonder model; in 12th place with 44 points. In preview on Google Agent Platform as `muse-wonder-1.0`.
Released on November 16; according to the Artificial Analysis model page, ranks 13th with 43 points, 9 points above Qwen3.8 Max (34). Available as an API in Model Studio under the name `qwen4-max`; 1M context, text, image, and video input. Singapore pricing is $1.50 / $4.50. Not open-weight.
Anthropic’s fast, low-cost model at $0.10 / $0.50. According to Meta, the new Muse Ember delivers the same score at about one-third of the task cost.
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.
The current model for the Grok app and API. Grok 4.8 is expected to become available to some Grok Bot users; it has no official API identifier, pricing, or Artificial Analysis score yet. Grok Bot 1.0 was released on November 11.