YZ·RADARyz.mikail.net
Daily edition · 9 November 2026

Tencent Yuanqi shuts down, Qoduck is free on Qoder; TOP 10 unchanged for a fifth day

For the fifth day, the TOP 10 rankings and scores remain unchanged. Gemini 4 Carbon leads with 60 points, followed by Claude Opus 5.5 (59) and GPT-6 Astra (57); the Artificial Analysis index is at v5.0.1. The latest accessible view of the LMArena text ranking is still dated October 27, and Gemini 4 Carbon is also in first place there. The watchlist is unchanged as well: Muse Wonder is 2 points short of the TOP 10 with 44 points. This report’s scan was conducted on the morning of November 9, and 42 new developments were added; 25 are from Tencent and 10 from Alibaba/Qwen.

The two most urgent items of the day are a shutdown and a breaking protocol change. Tencent’s agent platform Yuanqi discontinued all public services on November 9, including app access, data reading/writing, API, console, and published agent responses. Data can be downloaded through a separate export console until January 9, 2027; Tencent recommends that users who want to continue switch to ADP. Moonshot, meanwhile, removed the server_hello/client_hello and ack frames from Kimi Code’s KAP-server WebSocket protocol. Every command must now carry a `request_id` and receives a single response frame; plain `subscribe`/`unsubscribe` replace `subscribe_v2`/`unsubscribe_v2`. A command without a `request_id` closes the connection with code 1008. The change is on the main branch; users of external scripts or clients generated from AsyncAPI need to update.

Qoder stands out on pricing and access. The undisclosed model Qoduck is available to individual accounts on Qoder International (from Free to Ultra) with a daily free quota; calls within the quota use no Credits at a 0.0× rate, while 0.5× applies after the quota is exhausted. Teams/Enterprise accounts, as well as the IDE, JetBrains plugin, and CLI, are excluded from the promotion. On the same platform, Qwen3.8-Flash’s free period ended on November 7; the model now costs 0.1× during regular hours and 0.04× at a 60% discount from 14:00–00:00 UTC (17:00–03:00 TRT). OpenAI added typed APIs for scheduled automations (Automations) in the Agents API to Go SDK v3.97.0 and Ruby SDK v0.121.0: users can set up one-off schedules (`run_at`) or recurring schedules using a cron expression and IANA time zone. At Google, the `/go/` URL path for Gemini Notebook Enterprise began rolling out globally to locations outside us and eu on November 9; Cloud Asset Inventory can now also search Gemini for Google Cloud Topic resources.

It’s a busy day for Tencent’s developer tools. LoopForge administrators can now set a single shared CodeBuddy key for their team and restrict the allowed models, create accounts and reset passwords in the Member Accounts section; the platform image is also being released for linux/arm64. teamai-cli reads GitHub Copilot CLI sessions as read-only sources and writes WorkBuddy files to the data root of the most recently opened app. Tencent also released the DSpine draft model for Qwen3.8-27B (Qwen3.8-27B-DSpine-b16) and speculative inference support with SGLang. Response templates and generation tags were added to Qwen’s Hugging Face model cards; training on assistant outputs only (`assistant_only_loss`) is now easier with TRL. DeepSeek Harness opened an interest form for its official plugin marketplace and creator support program; the program has not started yet. No new changes were found for Anthropic, Meta, xAI, Z.ai, or Xiaomi during this period.

What to do today: If you have an agent or knowledge base on Yuanqi, back it up immediately through the export console and plan your move to ADP. If you connect to the Kimi Code server with your own WebSocket client, remove the handshake and switch to the `request_id` and `subscribe` contract. If you use Qoder, try Qoduck’s daily free quota and move long Qwen3.8-Flash jobs to after 17:00 TRT. Consider OpenAI Agents API Automations for recurring agent tasks.

Today’s TOP 10

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

  1. 01

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    At the top for a fifth day with 60 points. Access remains limited to select cyber defenders and Fairwind partners; general API and app access have not yet opened.Released: 28 October 2026

  2. 02

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Anthropic’s flagship model, available on Claude Pro/Max/Team/Enterprise and via API. No new changes from Anthropic during this period.Released: 22 September 2026

  3. 03

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

    OpenAI’s most powerful GPT-6 model, for long and difficult reasoning tasks. Scheduled automation support for the Agents API was added to Go SDK v3.97.0 and Ruby SDK v0.121.0.Released: 3 September 2026

  4. 04

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

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

  5. 05

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

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

  6. 06

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

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

  7. 07

    Gemini 4 Argon

    Google · Gemini=
    50pts

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

  8. 08

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta’s most powerful Muse model; no public API. Meta paused commercial terms updates that were due to take effect on November 20; the current terms remain in effect.Released: 12 October 2026

  9. 09

    Claude Fable 5.1

    Anthropic · Claude=
    50pts

    The model with the highest accuracy rate; at $10 / $50, it is the most expensive on the list.Released: 1 September 2026

  10. 10

    Kimi K3.1 Pro

    Moonshot AI · Kimi=
    46pts

    API at $2 / $10; open weights on Hugging Face (2.8T total / 104B active, 1M context). The handshake/ack were removed from the Kimi Code KAP-server WebSocket protocol; external clients need to switch to the `request_id` and `subscribe` contract.Released: 23 October 2026

Key signals

  • The TOP 10 rankings and scores remain unchanged for a fifth day: Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57. The watchlist is unchanged; 42 new developments were added today.
  • Tencent: The agent platform Yuanqi shut down on November 9. Data can be downloaded through the export console until January 9, 2027; Tencent recommends switching to ADP.
  • Moonshot: The handshake and ack were removed from the Kimi Code KAP-server WebSocket protocol; commands now carry `request_id`, and `subscribe` replaces `subscribe_v2`. External clients need to be updated.
  • Qwen/Qoder: The undisclosed model Qoduck is available to individual accounts with a daily free quota. Qwen3.8-Flash’s free period has ended; it is 60% off (0.04×) from 17:00–03:00 TRT.
  • OpenAI: Scheduled Agents API automations were added to Go SDK v3.97.0 and Ruby SDK v0.121.0. Google: The Gemini Notebook Enterprise `/go/` path is rolling out to other locations.
  • Tencent: LoopForge adds a shared CodeBuddy key, member accounts, and an arm64 image; teamai-cli reads Copilot CLI sessions. DeepSeek Harness opened an interest form for its plugin marketplace.

Worth trying today

01

Recover your Yuanqi data before January 9, 2027

Yuanqi shut down on November 9; apps, APIs, and published agents no longer respond. Sign in to the Yuanqi data export console and download agent settings, workflows, knowledge bases, conversation records, and custom tools separately; the console will remain open until January 9, 2027. If you plan to continue, follow Yuanqi’s ADP migration guide to create a new agent in ADP, import your knowledge bases and workflows, and verify its responses against the previous outputs.

Tencent · Hy
02

Move your Kimi Code WebSocket client to the new contract

If you connect to KAP-server with your own script, stop waiting for `server_hello` after connecting and remove `client_hello`/ack sends. Replace `subscribe_v2`/`unsubscribe_v2` calls with `subscribe`/`unsubscribe` commands in the form `{"request_id":"...","session_id":"...","agent_ids":[...]}`, and match responses using `response.request_id`. Reply with pong to the server’s `{ type: "ping" }` frame; treat a 1008 close as a missing `request_id` and code 40001 as an invalid subscription parameter. Keep Web, Desktop, and Remote Control clients on the same version family.

Moonshot AI · Kimi
03

Use Qoduck’s quota and Flash’s overnight discount on Qoder

Update Qoder to the latest version and select Qoduck in the model picker; no coupon or application is required, and calls within the daily quota use no Credits, with a 0.5× rate applying after the quota is exhausted. The promotion is limited to individual Qoder International accounts; it is not available in the IDE, JetBrains plugin, or CLI. Move long, multi-request jobs with Qwen3.8-Flash to 14:00–00:00 UTC, or 17:00–03:00 TRT: during these hours, the rate automatically drops to 0.04× from 0.1×. Note that if a feature calls another model in the background, that call is billed at its own rate.

Alibaba · Qwen
04

Schedule recurring agent tasks with OpenAI Automations

Upgrade the SDK with `go get github.com/openai/openai-go/v3@v3.97.0` for Go or `gem install openai -v 0.121.0` for Ruby. In Go, use `client.Beta.Agents.Automations.New` with `type: "periodic"`, `expression` (cron), and `timezone` (for example, `Europe/Istanbul`); for a one-off task, provide `type: "once"` and `run_at`. For tasks that need to preserve previous findings, target the existing session; for independent tasks, target a new session. You can pause and resume schedules using the same API and list their run records.

OpenAI · ChatGPT

This edition’s updates by AI

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

Google · Gemini

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

New in this edition · 2

  • AvailabilityGemini Notebook Enterprise

    Gemini Notebook Enterprise /go/ path rolling out to other locations

    According to the Gemini Enterprise release notes, the /go/ URL path is live in the global and us locations and rolling out in the eu location; rollout to all other locations begins on November 9, 2026. This is the availability phase of the previously announced URL change reaching new locations.

    Usage details

    Who can use it: Gemini Notebook Enterprise and Gemini Notebook in Gemini Enterprise; live in global and us, rolling out in eu, and available in other locations starting 2026-11-09.

    How to use it: Use the new path: notebooklm.cloud.google.com/go/REGION/ or, in Gemini Enterprise, vertexaisearch.cloud.google.com/notebooklm/go/global/. If your organization uses a URL allowlist or proxy, allow notebooklm.cloud.google.com/*, notebook.cloud.google.com/*, and the relevant vertexaisearch.cloud.google.com/notebooklm/* paths.

    Tip: Allowlisting the entire domain with /*, rather than individual paths, reduces the need to add a new rule when URL paths change in the future.

    ✓ Verified by official sourceSource: Gemini Enterprise sürüm notları

Anthropic · Claude

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

No notable change since the last edition.

OpenAI · ChatGPT

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

New in this edition · 3

  • APIOpenAI Go SDK

    Agent scheduling APIs added to the OpenAI Go SDK

    OpenAI Go SDK v3.97.0 added typed APIs for managing agent schedules and their execution records. Creating, listing, updating, pausing, resuming, and deleting one-time or recurring schedules is supported.

    Usage details

    Who can use it: Developers using OpenAI Go SDK v3.97.0 with access to the Agents API.

    How to use it: Upgrade the SDK with `go get github.com/openai/openai-go/v3@v3.97.0`; create a schedule through `client.Beta.Agents.Automations.New` using `type: "periodic"` with the `expression`/`timezone` fields, or `type: "once"` with the `run_at` field.

    Tip: Use an IANA time zone along with the cron expression to preserve the user’s local time; target the existing session instead of a new one if prior context is needed.

  • APIOpenAI Ruby SDK

    Agent scheduling APIs added to the OpenAI Ruby SDK

    OpenAI Ruby SDK v0.121.0 added typed APIs for managing agent schedules and scheduled runs. It supports creating, viewing, listing, updating, pausing, resuming, and deleting one-time and recurring schedules, as well as schedule runs.

    Usage details

    Who can use it: Developers using OpenAI Ruby SDK v0.121.0 with access to the Agents API.

    How to use it: Upgrade the SDK with `gem install openai -v 0.121.0`; use the Automations endpoints in the Agents API to define a schedule of type `once` or `periodic`, and provide a cron expression and IANA time zone for recurring jobs.

    Tip: To preserve a session’s historical findings, use its existing session ID as the schedule target; target a new session for independent jobs.

  • APIOpenAI Ruby SDK

    OpenAI Ruby SDK added gpt-6-rosalind to beta Responses requests

    OpenAI Ruby SDK v0.121.0 added the `gpt-6-rosalind` model to the known model values for beta Responses creation, compaction, and client event requests. Separately from earlier model catalog support, this provides typed model support for these beta request surfaces in the Ruby SDK.

    Usage details

    Who can use it: Only internal research organizations and users approved in OpenAI’s trusted-access program; API access is also required.

    How to use it: Upgrade the SDK with `gem install openai -v 0.121.0` and use `gpt-6-rosalind` in the beta Responses request’s `model` field; the request will not work unless your access is approved under trusted-access.

    Tip: Keep the model ID configurable through application settings; research access and pricing are subject to API terms separate from a standard ChatGPT subscription.

Moonshot AI · Kimi

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

New in this edition · 1

  • APIKimi Code KAP-server

    Handshake and ack frames removed from the KAP WebSocket protocol

    KAP-server no longer sends a server_hello/client_hello and ack sequence when a connection is established; all commands now include a request_id and receive a single response frame. Plain subscribe/unsubscribe are used instead of subscribe_v2/unsubscribe_v2; ping content is now only { type: "ping" }, the invalid subscription code was moved to 40001, and commands received without a request_id are closed with 1008.

    Usage details

    Who can use it: The KAP-server change merged into the Kimi Code main branch on 2026-11-09; it is prepared for matching Kimi Code Web/Desktop/Remote Control clients. External WebSocket scripts using the old protocol and integrations that generate clients from AsyncAPI must migrate to the new contract; there is not yet a separate entry in the published CLI/Desktop release notes.

    How to use it: In your external WebSocket client, remove the initial wait for server_hello and the sending of client_hello/ack; convert subscribe_v2/unsubscribe_v2 calls to subscribe/unsubscribe commands in the {"request_id":"...","session_id":"...","agent_ids":[...]} format. Match responses using response.request_id, reply to the server's plain ping frame with pong, and handle a missing request_id as a 1008 connection closure.

    Tip: First update the client according to the current Kimi Code server-api.md contract, and keep Web/Desktop/Remote Control and external scripts in the same version family; mark 40001 as a permanent subscription parameter error.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 25

  • SettingLoopForge

    LoopForge adds a shared CodeBuddy key for teams

    Administrators can now save a single CodeBuddy API Key in a write-only field and specify the network environment and which models can be used. Members without their own key can use this key to chat and run SOPs only with permitted models; members with their own key keep their catalog and default settings.

    Usage details

    Who can use it: Administrators on the LoopForge platform and team members without their own CodeBuddy key.

    How to use it: Open Platform Management → Shared CodeBuddy Key, then save the key, network environment, and list of permitted models. Members will then use the shared key when chatting or running SOPs.

    Tip: Start with the least-privileged model list for the shared key, and store it in a secure password manager, assuming the write-only key field will not show it again.

  • RetiredYuanqi

    Yuanqi platform discontinues its public services

    On November 9, 2026, Yuanqi is discontinuing all public services, including app access, data reading and writing, APIs, the console, and responses from published agents. Users can back up their data through the export console until January 9, 2027; Tencent recommends switching to ADP and provides related user rights.

    Usage details

    Who can use it: All Yuanqi users; the active platform shuts down on November 9, while the data export console remains open until January 9, 2027.

    How to use it: Use Yuanqi's export tool before shutdown; after shutdown, sign in to the Yuanqi data export console and download your data before January 9, 2027. To continue, create a new agent in ADP and import your knowledge base and workflows.

    Tip: Export workflows, knowledge bases, conversation records, and custom tools separately—not just agent settings—and revalidate them in ADP.

  • AvailabilityLoopForge

    LoopForge platform images released for linux/arm64

    The LoopForge platform image is now available for linux/arm64 alongside linux/amd64. Apple silicon Macs and ARM servers can now pull the native ARM image; CodeBuddy is said to run more slowly under QEMU emulation on x86, and the model list cannot be read.

    Usage details

    Who can use it: Native linux/arm64 for Apple silicon Macs and ARM servers; emulation on x86 is not recommended.

    How to use it: Pull and run the LoopForge platform image from GHCR on a linux/arm64 host using the target version tag; on an x86 host, choose the linux/amd64 image instead of emulating it.

    Tip: Use the native image on ARM servers; choose the amd64 image to avoid slowdowns and CodeBuddy model list issues under QEMU.

  • FeatureLoopForge

    LoopForge adds member accounts and password resets to platform management

    Platform administrators can now list accounts, create password-based accounts with optional administrator privileges, and reset passwords from Platform Management. Adding passwords to GitHub or gateway accounts is not allowed.

    Usage details

    Who can use it: LoopForge platform administrators; actions require administrator privileges, the same origin, or an API token.

    How to use it: Open Platform Management → Member Accounts; use the account creation form to select a user and, if needed, the administrator option. Reset an existing account from the same section.

    Tip: For teams with self-registration disabled, create at least one platform administrator account first and test the recovery flow.

  • APIteamai-cli

    teamai-cli aligns the workflow output schema with admission

    Workflow output now declares the definition and optional workflows fields that admission actually accepts; invalid scorers and teamRevision fields have been removed from the schema. The expected output schema and the runtime strict BundleSchema are now derived from a single source.

    Usage details

    Who can use it: Steps that produce teamai workflow output and agents that use the workflow type as JSON Schema.

    How to use it: When using the workflow type in a step output, produce only the definition and, if needed, workflows fields; do not send scorers or teamRevision.

    Tip: For custom workflow agents, use the expected output contract generated by teamai rather than copying the schema manually.

  • Capabilityteamai-cli

    teamai-cli sends WorkBuddy files to the root of the most recently launched app

    WorkBuddy user-scoped skill, rule, team instruction, agent, and MCP files are now written to the .workbuddy or .workbuddy-ai root based on the last-launch.json record identifying the most recently launched app, rather than only on which directory exists. Project scope continues to carry only skills; settings.json and hooks remain in .workbuddy.

    Usage details

    Who can use it: teamai-cli pull/doctor in environments with WorkBuddy and WorkBuddy AI installed; systems where both data roots may exist.

    How to use it: Launch WorkBuddy or WorkBuddy AI once to create the correct app’s last-launch.json record; then run teamai pull and teamai doctor to verify the target root.

    Tip: An old data root may remain after an app is uninstalled, so don’t clean up based only on whether a directory exists; check last-launch.json and the doctor output first.

  • CapabilityBrowserSkill

    BrowserSkill preserves upload error receipts when the debugger detaches

    The file upload handler preserves its own error receipt even if it has to detach the debugger connection. This prevents reasons such as set_file_input_failed or file_drop_failed from being lost; resynchronization is required before the next page operation.

    Usage details

    Who can use it: tool.upload in the BrowserSkill extension; file input and file drop flows that experience a debugger detach.

    How to use it: Try uploading a file with tool.upload; if it returns an error, read the reason, phase, and effect_state fields, sync the page, then try again.

    Tip: Instead of blindly retrying automatically after an error, first process the effect_state and recovery instructions in the receipt; assume the file may have been partially applied.

  • CapabilityBrowserSkill

    BrowserSkill extract applies CSS case transformations

    extract now applies CSS uppercase/lowercase transformations to table cells, headings, list fields, and discovery previews. Nested style overrides, language information, and the text byte budget are preserved; HTML attribute reads are unchanged.

    Usage details

    Who can use it: Table, list, and discovery flows that use the BrowserSkill extract tool.

    How to use it: Run an extract call with a table, list, or discover action; compare the text values with text-transform: uppercase/lowercase applied against the separate attribute fields.

    Tip: If you need to distinguish the displayed value from the raw attribute value, define text and attribute reads as separate fields in list fields.

  • CapabilityBrowserSkill

    BrowserSkill fill accepts normalized end-of-editor caret positions

    fill now treats end-of-editor caret positions normalized by the editor as equivalent, even if the caret has moved to the last paragraph or text node. Selections that actually leave the editor or lose focus are still rejected.

    Usage details

    Who can use it: The fill tool in the BrowserSkill extension; rich-text and contenteditable editors.

    How to use it: Before calling fill, focus the contenteditable editor and move the caret to the end; verify that text is inserted when the editor normalizes the caret to a child node.

    Tip: For applications that move the caret on input or selectionchange events, test focus loss and genuinely out-of-bounds positions separately.

  • Featureteamai-cli

    teamai-cli skips team folders it cannot read and displays them in the dashboard

    pull now skips entire folders containing relevant marker files under workflows/, hooks/, rules/, mcp/, benchmarks/, scorers/, and agents/, while continuing to pull the remaining resources. Skipped folders are listed in pull, status, doctor, review, and the dashboard with their type, marker, and update command.

    Usage details

    Who can use it: teamai-cli users; team repositories containing marker-based folders that this version cannot yet read.

    How to use it: Run teamai pull; open the Skipped folders section in the dashboard. Once support is available, run teamai update && teamai pull.

    Tip: Update all team members' teamai versions before adding marker-based folders to the team repository; older versions may read the files as the wrong type.

  • Capabilityteamai-cli

    teamai-cli made concurrent writes to the review queue safe

    The import, review --apply, and review --reject operations that write to .teamai/pending-review.jsonl now use a shared lock. Concurrent additions or deletions no longer drop sections added by another operation; if the lock is busy, the command fails without changing the queue.

    Usage details

    Who can use it: Teams running teamai review and import flows concurrently in the same checkout.

    How to use it: Run teamai review --apply or --reject; if you get the "Another teamai command is changing" error, retry the same command after the other write operation finishes.

    Tip: Serialize jobs in CI and hooks that write to the same pending-review file in parallel; the new lock reports conflicts explicitly instead of hiding them.

  • Plugin / Integrationteamai-cli

    teamai-cli sessions can now read Copilot CLI transcripts

    A copilot platform adapter has been added to the session flow. It reads GitHub Copilot CLI's session-state event logs under ~/.copilot as a read-only source. Because Copilot resumes from its own logs, teamai lists these sessions as sources, not as writable targets for another tool.

    Usage details

    Who can use it: teamai-cli session users in environments with GitHub Copilot CLI installed; Copilot transcripts are read-only.

    How to use it: Check for copilot support with teamai session platforms; then view the ~/.copilot logs with teamai session list or the relevant session command.

    Tip: When moving Copilot sessions to another tool, plan to use them as a source for archiving and context transfer, not as resumable targets.

  • Capabilityteamai-cli

    teamai-cli added blocking timeouts to team repository remote calls

    The fetch, push, and ls-remote calls used by teamai status, recall promote, and push now run with blocking timeouts and progress reporting. An unresponsive SSH remote command now fails or returns the branch status as unknown instead of holding a lock indefinitely.

    Usage details

    Who can use it: Teams using teamai status, recall promote, and push with SSH or slow remote team repositories.

    How to use it: Use the usual teamai status or teamai recall promote commands; if a connection hangs, catch the timeout error, fix the remote, and rerun the command.

    Tip: To avoid bringing back indefinite waits, verify that your own SSH wrapper also applies timeouts to teamai's remote calls.

  • CapabilityBrowserSkill

    BrowserSkill uses computed visibility for hidden content

    extract now distinguishes when the hidden attribute is hidden only by the browser's default CSS rule. If author CSS restores the display value, the content is extracted; actual hidden=until-found or overlay states remain hidden.

    Usage details

    Who can use it: BrowserSkill extract; pages that use the hidden attribute and author CSS.

    How to use it: Use a table or parent element with the hidden attribute as an extract target, but with display restored by CSS; verify that visible text appears in the output.

    Tip: Do not filter automation targets based only on the hidden attribute; also test actual computed visibility and content-visibility behavior.

  • APIteamai-cli

    teamai-cli added integer validation for --port and --project-id

    dashboard --port now requires a safe integer in the 0–65535 range, while bind-project --project-id requires an integer of at least 1. Invalid, fractional, or NaN-converting inputs now produce readable errors instead of being passed through silently.

    Usage details

    Who can use it: Scripts and CI jobs that use the teamai-cli dashboard and bind-project commands.

    How to use it: Use valid values such as teamai dashboard --port 3721 and teamai bind-project --project-id 42; catch the returned error for invalid values and fix the command.

    Tip: In your CLI wrappers, check the new error code instead of leaving options to Number; do not treat 1.5, 1e3, 0x10, or out-of-range values as valid.

  • Capabilityteamai-cli

    teamai-cli translated ClawPro project-binding prompts into English

    ClawPro local-agent project-binding prompts, options, and hook context are now displayed in English. Commands and project IDs are unchanged; only the interaction text and guidance sent to the agent were translated.

    Usage details

    Who can use it: English-language interfaces and agents that use ClawPro project binding in teamai-cli.

    How to use it: Start the CodeBuddy/ClawPro flow in an unbound workspace or run teamai bind-project; use the Available projects and Bind to a project? prompts.

    Tip: If your automations parse prompts, rely on the project ID and teamai bind-project command instead of fixed Chinese text.

  • APIteamai-cli

    teamai-cli rejects numeric boundary options for import

    import --max-repos must now be a positive integer; --depth in repo import is also validated as a positive integer. Invalid values now produce an error and stop before writing, instead of silently proceeding with NaN or an invalid limit.

    Usage details

    Who can use it: Users of teamai-cli import --from-org and repo import flows.

    How to use it: Use a positive integer, such as teamai import --from-org acme --max-repos 25; for --depth in repo import, provide a value of 1 or greater.

    Tip: Validate --max-repos and --depth as positive integers in CI before running the command; otherwise, you will now get a failed command instead of zero results.

  • Settingteamai-cli

    teamai-cli reports automatic update outcomes with separate statuses

    The update handler records completed operations as ok, operations that were unnecessary or not permitted as skipped, and check/install/hook refresh errors as failed. It leaves installations without automatic hook support at the debug level and displays instructions for manual updates.

    Usage details

    Who can use it: Installations using teamai update and automatic hook updates.

    How to use it: Run teamai update or wait for the automatic hook to trigger; distinguish the result as ok, skipped, or failed in the diagnostics output and debug log.

    Tip: In your monitoring system, track these three statuses as separate metrics rather than monitoring only the command’s exit code; skipped does not mean failure.

  • CapabilityBrowserSkill

    BrowserSkill records clicks on controls in non-editable editor regions

    The recording flow now preserves buttons and links in contenteditable="false" regions inside an editor as click steps instead of swallowing them as editor text. As a result, clicks on actual controls such as attachment cards are not lost.

    Usage details

    Who can use it: Pages using the BrowserSkill extension’s recording feature, especially those with contenteditable editors.

    How to use it: Start a recording with BrowserSkill and click a button in a contenteditable="false" area inside the editor; check that a separate click step appears in the recording output.

    Tip: Include attachment or embedded-card controls in rich-text editor test scenarios, and verify the resulting target role and name.

  • CapabilityBrowserSkill

    BrowserSkill records changes to native selects in open Shadow DOM

    The recording flow now captures native select controls in open Shadow DOM, where the change event does not reach document, through the input event. The existing path for Light DOM selects is preserved while duplicate recordings are prevented.

    Usage details

    Who can use it: BrowserSkill recording feature; native select controls in open Shadow DOM.

    How to use it: Start a recording, change the value of a native select inside an open shadow root, and check that the output contains a single select step with value and label information.

    Tip: In web component tests, don’t rely only on the document-level change event for select changes; also verify the recorded step.

  • CapabilityBrowserSkill

    BrowserSkill press sends physical codes for punctuation keys

    press now generates physical code and virtual key code values for ASCII punctuation characters on the US keyboard layout. This ensures shortcuts such as Ctrl+/ reach page handlers that check event.code.

    Usage details

    Who can use it: Automations using the press interaction tool in the BrowserSkill extension.

    How to use it: Call the press tool with a shortcut such as Ctrl+/, Ctrl+-, or Ctrl+= and verify that it triggers the target page’s corresponding event.code handler.

    Tip: When testing keyboard shortcuts, check not only the visible character but also the physical code value expected by the page.

  • CapabilityDSpine

    DSpine added support for Qwen3.8-27B inference and SPS tuning

    DSpine added a dynamically verified inference flow for Qwen3.8-27B, ReplaySSM integration, and an SPS tuning tool. The README also provides an example using Python 3.11+, SGLang 0.5.22, and up to eight concurrent requests on a single GPU.

    Usage details

    Who can use it: CUDA/SGLang users of the DSpine GitHub repository; the new flow requires Python 3.11+ and SGLang 0.5.22.

    How to use it: Install DSpine; use the --speculative-algorithm DSPINE, --speculative-draft-model-path tencent/Qwen3.8-27B-DSpine-b16, and --enable-linear-replayssm-spec options in the sglang serve command. Run scripts/tune_sps.py for your workload and pass the resulting file with --speculative-dspark-sps-table-path.

    Tip: Regenerate the SPS profile when the GPU, model, runtime, or workload changes; the profile depends on the hardware and request distribution.

    ✓ Verified by official sourceModels: Qwen3.8-27B, Qwen3.8-27B-DSpine-b16Source: DSpine resmi commit: Qwen3.8-27B inference desteği
  • ModelDSpine

    DSpine draft model released for Qwen3.8-27B

    Tencent released the five-layer Qwen3.8-27B-DSpine-b16 draft checkpoint for Qwen3.8-27B on Hugging Face. The model card documents a block size of 16, a sliding window of 8.192 tokens, and BF16/SGLang usage.

    Usage details

    Who can use it: Publicly available on Hugging Face; for users of Qwen3.8-27B with DSpine and SGLang.

    How to use it: Follow the usage steps in the model card; install the DSpine repository and use --speculative-draft-model-path tencent/Qwen3.8-27B-DSpine-b16 in the SGLang command.

    Tip: Tune the SPS profile with your own GPU and workload before production traffic; don’t generalize the benchmark values in the model card directly.

    ✓ Verified by official sourceModels: Qwen3.8-27B, Qwen3.8-27B-DSpine-b16Source: Hugging Face model kartı: Qwen3.8-27B-DSpine-b16Source: Hugging Face commit geçmişi
  • Capabilityteamai-cli

    teamai-cli preserves delivery records across HOME aliases

    pull no longer changes the checkout record because of a trailing separator in the HOME path, a symbolic link, or platform-specific syntax. Hashes of files delivered in earlier records are carried over to the new canonical path, preventing edited user files from being accidentally deleted.

    Usage details

    Who can use it: teamai-cli installations with differently written HOME variables, symbolic links, or Windows path formats.

    How to use it: Run teamai pull after using the same teamai checkout with different HOME formats; check the doctor and pull output to confirm that the delivery record is preserved.

    Tip: On machines that move HOME through a symbolic link, don’t manually delete the old state file; first run a pull with the new version to migrate the delivery ledger.

  • Capabilityteamai-cli

    teamai-cli correctly reads YAML dates during learning maintenance

    recall maintenance and promotion now handle quoted, unquoted, or timestamped YAML formats in the frontmatter date field the same way. The time information in timestamps is preserved when applying age thresholds.

    Usage details

    Who can use it: Knowledge/learnings repositories that use teamai recall maintenance and teamai recall promote.

    How to use it: Use date: 2026-11-07, date: "2026-11-07", or an ISO timestamp in the learning frontmatter; then run teamai recall maintenance --update-quality or teamai recall promote.

    Tip: For automations that depend on age thresholds, write dates in ISO format and, if possible, specify the time zone explicitly.

Alibaba · Qwen

Qwen3.8 Max
Full history (673) →

New in this edition · 10

  • ModelQoder

    Daily free usage begins for Qoduck, a hidden model in Qoder

    Qoder International has made the as-yet-unidentified Qoduck model available to new and existing individual users with a limited-time daily free quota. Calls within the daily quota run at a 0.0× rate and consume no Credits; the standard 0.5× rate applies once the quota is used up.

    Usage details

    Who can use it: Individual Free, Pro Trial, Lite, Pro, Pro+ and Ultra accounts on Qoder International are eligible; Teams and Enterprise accounts cannot see Qoduck. The promotion is not supported in Qoder IDE, the JetBrains plugin, Qoder CLI or other products.

    How to use it: Update Qoder to the latest version, select Qoduck from the model picker and start a task. No request or coupon is required.

    Tip: Paid plans get a higher daily free quota; if you use it up, wait for it to renew the next day or temporarily accept the standard 0.5× rate.

    ✓ Verified by official sourceModels: QoduckSource: Qoder: Qoduck için sınırlı süreli ücretsiz kullanım
  • ModelQwen

    Transformers and TRL compatibility added to Qwen model cards

    In Qwen’s Hugging Face model repositories, the response template can split model output into reasoning, content, and tool calls fields, then reconstruct it through the Transformers parse_response flow, while remaining compatible with the vLLM parser. Generation tags in the chat template let TRL assistant_only_loss and return_assistant_tokens_mask flows train only on assistant output; the tags do not change the generated text.

    Usage details

    Who can use it: The change has been published on the relevant main branches of the Qwen3.6 and Qwen3.8 model cards; this is a chat template and tokenizer metadata update, not a change to the weights. It affects Transformers, TRL, or vLLM users who fetch the latest files from Hugging Face.

    How to use it: Re-download the relevant model from Hugging Face and use the updated chat template with AutoTokenizer or AutoProcessor.apply_chat_template. For training, use return_assistant_tokens_mask=True and the TRL assistant_only_loss flow; for parsing, use Transformers parse_response or a compatible vLLM parser.

    Tip: During fine-tuning, use up-to-date versions of Transformers and TRL that both support generation tags to include only assistant tokens in the loss.

    ✓ Verified by official sourceModels: Qwen3.6-27B, Qwen3.6-27B-FP8, Qwen3.6-35B-A3B, Qwen3.6-35B-A3B-FP8, Qwen3.8-27B, Qwen3.8-27B-FP8, Qwen3.8-Flash-Next, Qwen3.8-Flash-Next-FP8, Qwen3.8-2.4T-A95B, Qwen3.8-2.4T-A95B-FP8Source: Qwen Hugging Face etkinlik akışıSource: Qwen3.6-27B tartışması: response template ve generation tagsSource: Qwen3.8-Flash-Next tartışması: response template ve generation tags
  • Pricing / PlanQoder

    Qwen3.8-Flash free promotion ends and off-peak discount begins

    The daily free usage period for Qwen3.8-Flash on Qoder International ended at 00:00 on November 7, 2026. The model is now billed at 0.1× during regular hours and at 0.04× during off-peak hours, a 60 percent discount.

    Usage details

    Who can use it: Available in Qoder, Qoder IDE, the JetBrains plugin, Qoder CLI, QoderWake, Cloud Agents, Mobile and the web app. The off-peak discount is applied automatically from 14:00–00:00 UTC; in Türkiye, this corresponds to 17:00–03:00.

    How to use it: Update the Qoder product, select Qwen3.8-Flash from the model picker and send your request between 14:00 and 00:00 UTC. No separate registration or setting is required for the discount.

    Tip: Move long-running tasks with many requests to off-peak hours; if a feature calls another model during execution, that call may be billed at that model’s own rate.

  • CapabilityQwen Code

    Auto-mode records the first commit in an empty Git repository as an amend exemption

    When a command elevated with Ctrl+B creates the first commit in a Git repository that has none yet, Qwen Code now correctly records that commit for the Auto-mode amend exemption. This prevents subsequent automatic changes from triggering incorrect protection behavior because of the first commit.

    Usage details

    Who can use it: Applies to users running Qwen Code v0.27.0 in Git repositories that have not yet had their first commit created with Auto-mode.

    How to use it: Run Qwen Code in Auto-mode in a Git repository with no commits, and approve with Ctrl+B when an elevated command is required; the first commit is automatically recorded as an exemption.

    Tip: Before the first commit, check which files Auto-mode may touch, and review the working tree after creating the first commit.

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

    CI Failure Patrol publishes visible bilingual notices during retries

    When Qwen Code’s CI Failure Patrol workflow retries flaky test failures or cannot classify a failure, it publishes a visible bilingual notice. This lets pull request contributors see directly whether a failure was retried or could not be classified.

    Usage details

    Who can use it: Applies to CI Failure Patrol pull request checks in the Qwen Code GitHub repository and contributors who view their results.

    How to use it: Open a pull request in the Qwen Code repository or open the CI Failure Patrol check result for an existing pull request; the notice appears automatically in the check output.

    Tip: Before starting a new run, check whether the notice indicates a retry for a flaky failure or an unclassified failure.

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

    Experimental unifiedAgentLoop continues Ink TUI turns after tools run

    When experimental.unifiedAgentLoop is enabled, Ink TUI sends a continuation turn after a tool group and manages queued inputs and retries within the same turn flow. This reduces interruptions after tool calls in interactive terminal sessions using the experimental shared agent loop.

    Usage details

    Who can use it: Applies to users of Ink TUI in Qwen Code v0.27.0 who enable the experimental.unifiedAgentLoop setting.

    How to use it: Set experimental.unifiedAgentLoop to true in Qwen Code settings, restart Ink TUI and run a turn that uses tools.

    Tip: If you encounter problems with the experimental flow, turn the setting off and return to the standard loop; observe the results especially when multiple user inputs are queued.

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

    File saves complete even when temporary file permissions are denied

    When some file systems reject changes to temporary file permissions, Cursor saves no longer fail before publishing. Qwen Code continues saving the file in this case.

    Usage details

    Who can use it: Applies in Qwen Code v0.27.0 to workspaces where the file system blocks temporary file permission changes.

    How to use it: Update Qwen Code to v0.27.0 and use the normal file editing and saving workflow; no separate setting or command is required.

    Tip: Check the file’s actual contents and Git diff after saving; this is especially useful on network drives or file systems with restricted permissions.

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

    Managed Web Shell pane and owner navigation reliability improves

    Issues affecting shell panes, host status updates, session measurement and owner navigation flows in Managed-host Web Shell have been fixed. This reduces inconsistent UI states when monitoring a remote Managed Agent session and opening its associated owner.

    Usage details

    Who can use it: Qwen Code v0.27.0 Managed Agent Web Shell sessions, especially the shell and session view running on a managed host.

    How to use it: Update Qwen Code to v0.27.0 and open a Managed Agent Web Shell session; the feature applies automatically and requires no additional settings.

    Tip: If the host status or owner link looks out of date, check the session’s actual running status before refreshing the browser; try reopening it in the latest version.

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

    Web Shell correctly handles transport capabilities and user-echo timing

    Web Shell now takes optional transport capabilities into account and avoids making unnecessary assumptions for transports that do not provide them. Local prompt timing is also correctly preserved while waiting for an authoritative user echo.

    Usage details

    Who can use it: Applies to Qwen Code v0.27.0 Web Shell and transport layers that use optional transport capabilities.

    How to use it: Upgrade Qwen Code to v0.27.0 and send a message through Web Shell; transport capabilities are detected automatically, with no manual configuration required.

    Tip: If you’re developing a custom transport layer, explicitly declare the capabilities it supports and don’t treat a user echo as confirmed until the authoritative event arrives.

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

    Channel turn completes if the response is delivered, despite a follow-up notification error

    If a response has been delivered to the user during a channel turn, the turn is no longer considered failed even if only the follow-up notification fails to send. A turn is marked as failed only if no response was delivered.

    Usage details

    Who can use it: Applies to Qwen Code channel sessions and integrations that use follow-up notifications for those sessions.

    How to use it: Update to Qwen Code v0.27.0 and run a task through a channel; the turn will appear as completed in the history even if a follow-up notification error occurs after the response is delivered.

    Tip: Check whether the main response was delivered before starting the same task again; rerunning it when only the follow-up notification is missing may create a duplicate operation.

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

DeepSeek · DeepSeek

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

New in this edition · 1

  • Plugin / IntegrationDeepSeek Harness

    Applications of interest open for the official plugin marketplace and author support program

    DeepSeek Harness announced plans to establish an official plugin marketplace and improve plugin API stability, compatibility, and security. A support program for plugin authors, including token grants, technical consultations, and promotion, is in the works; applications of interest are currently the only submissions being accepted, and launch rules for the marketplace and program have not yet been published.

    Usage details

    Who can use it: Planned for DeepSeek Harness plugin authors and community developers, as well as public preview users worldwide. The interest form is open, but the official marketplace and author support program are not yet available.

    How to use it: You can install a community plugin package by entering its name on the plugin installation page in the Plugins section of DeepSeek Harness. To join the author support program, fill out the interest form in the official DeepSeek Harness announcement; publishing rules will be announced separately when the marketplace opens.

    Tip: Clearly specify the plugin package name, repository link, target workflow, and the technical or security support you need in your application; test plugin compatibility against current API behavior before the marketplace opens.

Watchlist

Muse Wonder

44
Meta · Meta AI

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

Claude Haiku 5.5

36
Anthropic · Claude

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

Grok 4.7

35
xAI · Grok

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

Tencent Hy4

35
Tencent · Hy

Available via API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, free on WorkBuddy and CodeBuddy through November 15. Tencent’s agent platform Yuanqi shut down on November 9; a move to ADP is recommended.

Qwen3.8 Max

34
Alibaba · Qwen

Alibaba’s cloud flagship model; Qwen 4 has not yet been released. Qwen3.8-Flash’s free period on Qoder has ended and an overnight discount has begun; the undisclosed Qoduck model is available with a daily free quota.

DeepSeek V4.1 Flash 1017

33
DeepSeek · DeepSeek

MIT-licensed open weights, 1M context, 33 points. DeepSeek Harness opened an interest form for its official plugin marketplace and creator support program.

Sources