Kimi Code 2.5.2
The publisher released version 2.5.2 of Kimi Code on npm.
Model launches, product changes and release notes from the frontier labs, day by day.
Coverage through
134 updates · 7 labs · 30 days
134 updates in this edition
Times are Eastern. “Date only” means the publisher gave a day without a time.
Selected official release feeds refresh hourly. The editorial review of lab announcements and changelogs covers releases through .
No update recorded from the automatically checked feeds for this day.
No update recorded from the automatically checked feeds for this day.
The publisher released version 2.5.2 of Kimi Code on npm.
gemini-embedding-2-preview is deprecated. Effective November 6, 2026, API requests to gemini-embedding-2-preview route to gemini-embedding-2 at Gemini Embedding 2 rates. Update your model string to gemini-embedding-2. gemini-embedding-2…
Eligible Enterprise organizations can now tag projects with the codes they already use, like a charge code or client-matter number, and see most spend in those projects by tag. This feature is available in beta. For more information, see…
Release note: “Open the agent command center with codex agents --no-daemon. The command center and session picker now mark unread results”
Release note: “Changed auto mode with telemetry or feature flags off, including behind LLM gateways: calls the server can't check are asked…”
The publisher released version 2.5.1 of Kimi Code on npm.
The Gemini API now accepts PowerPoint 97-2003 (PPT), OpenDocument Presentation (ODP), and OpenDocument Spreadsheet (ODS) files as input, and adds page-image processing for Word 97-2003 (DOC) files. DOC, PPT, and ODP files are rendered as…
Effective December 3, 2026, API requests to gemini-omni-1.1-flash will automatically route to gemini-omni-1.2-flash at Gemini Omni 1.2 Flash rates.
Released Gemini Omni 1.2 Flash (gemini-omni-1.2-flash), featuring up to 15-second video generation and conversational editing (3 to 10 seconds for first and last frame interpolation, and 3 to 10 seconds per-turn video extension up to 30…
Tools can now return Office and OpenDocument files, audio, and video to the model. GenerateContent API: Function response parts now accept audio, video, and Office and OpenDocument files, in addition to images, PDFs, and plain text. They…
Claude Enterprise has new security and admin features, including a security center, inference hooks, centralized management, and additional access controls, several in beta. For more information, see our blog post: More visibility and…
We’re rolling out Personal Finance, a beta that lets you link your bank, credit card, loan, and investment accounts to Claude, so you can see your balances and spending on one page and ask Claude questions about your own money. It's…
Use GPT-6-Cyber to investigate vulnerabilities, reproduce them in an authorized test environment, and verify fixes. See Daybreak setup before sending your first request.
You can now set up an IP allowlist for your organization in the Claude Console, in beta. Requests to covered services from addresses not on your list are refused, which limits where a leaked API key can be used. Add IPv4 and IPv6…
You can now sign a member of your Claude Enterprise (claude.ai) organization out of Claude everywhere with the Admin API. This capability is in beta. See the Admin API reference. The member stays in the organization and can sign in again.…
Release note: “Changelog for v0.65.0”
The publisher released version 2.5.0 of Kimi Code on npm.
Released GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare for image generation and editing, with adjustable complexity, SVG generation, and animated GIF generation. Added markup-guided editing for supported GPT Image 2.5 and 2.6 models. See…
Released Voice remixing, which lets you modify an existing prebuilt or designed voice using natural-language instructions to create a new, independent stored voice. Use POST /v1beta/voices with gemini-3.8-flash-tts (or remix interactively…
Anthropic published this stable package version. The GitHub release list and public changelog did not identify individual changes.
Text from Mythos 5, Opus 4.7 and Sonnet 4.6 carries a watermark on supported platforms. Some non-default sampling settings on Sonnet 4.6 can prevent it.
The Sites plugin can deploy a website. This release also adds Kimi K3.1 Pro configuration and restarts paused workflows when a quota resets.
The scheduler flags connectors it added without a request, and network calls that expose the account email now prompt for confirmation.
Google removed both models from the AI Studio selector. On December 1, API requests will route to 3.8 Flash for Pro and 3.5 Flash-Lite for Flash.
Requests to grok-imagine-image-quality now route to Imagine Image 2.0 with the low quality setting.
The Messages API can keep a short-lived conversation on the server so follow-up requests send only the new turns.
No dated update recorded from the tracked sources.
October 31 is the read-only milestone in the published Evals migration schedule. The dashboard and API shutdown are scheduled for November 30.
Existing accounts retain access for now. Moonshot plans to retire the series by the end of December.
Anthropic published this stable package version. The GitHub release list and public changelog did not identify individual changes.
A single read now stops at 100,000 characters or roughly 25,000 tokens. Developers can raise the limit in their configuration.
Z.ai published Intel and ARM installers for macOS and Windows, along with checksums for verifying downloads.
The company describes earlier interventions in ChatGPT and additional safeguards for younger users alongside a new safety report.
Developers can give Gemini the Computer Use tool through either the Interactions API or the GenerateContent API.
Authorized researchers can use the biology-focused model in Work, Codex and the API. Existing Discovery approval does not confer Research access.
Anthropic plans to remove fast mode from Opus 4.8 on November 30. Applications that need fast mode can move to Opus 5.5.
The iOS and Android rollout lets people stay signed in to two accounts. Each retains its own chats, memory and subscription.
A permission setting stops Write, Edit and NotebookEdit from changing files outside approved working directories.
Terminal users can search conversation history, copy a selection and diagnose network failures with codex doctor --network-only.
The change targets Free, Go and signed-out users, plus the fallback model used after some paid users reach their message limit.
Pro 100 and Pro 500 allowances double. Pro 200 subscribers on the lower allowance move to the higher limit.
The revision updates guidance on intent, oversight, personalization and the effects an agent’s actions can have outside a conversation.
Console and Enterprise admins can break down token usage and costs by workspace, model, key or requester. The older reporting endpoints remain supported.
Import the previous 15 or 30 days of sessions. The CLI can also refresh a provider’s model list while running and reconnect dropped MCP servers.
Users can choose a supported thinking level explicitly. Specialized medical routing remains an exception.
Qualifying Pro and Business customers can link an API organization to activate promotional credits, separate from their ChatGPT usage allowance.
The Messages and Token Counting APIs now accept requests up to 128 MB, four times the previous limit.
A beta for Pro users in the US lets their dot exchange texts with them and send proactive updates using its existing memory and connections.
The CLI synchronizes output, isolates spinner rendering and addresses flicker and border glitches in tool output.
Admins can map users using an OIDC claim such as Microsoft Entra’s object ID instead of relying on the subject field.
The CLI can use Google Cloud Vertex AI and filter saved sessions by repository. Supported models can also change reasoning effort during a conversation.
The new connections let users look up messages, check their calendar and update events on supported clients.
The gemini-api command generates text or media, analyzes files and runs managed agents from scripts or CI without maintaining session state.
Developers can set how text responses handle profanity, violence and sexual content through Chat Completions or the Responses API.
Version 1.2026.293 adds inline playback and lets you open a subagent’s conversation from task activity.
Hosts can tell an idle runner that a process it started is still active and set a maximum hold time.
The CLI also stops displaying a failed swarm when it rejected the request before starting any agents.
Existing unmigrated GPTs remain usable until their retirement date. OpenAI points teams to Plugin Creator for new reusable workflows.
The new model supports visual reasoning and computer use. Eligible ChatGPT customers can select it in Work or Codex.
New paid projects retain requests unless developers change the project setting or send store=false. Existing projects keep their previous configuration.
Interactive users enable swarm mode explicitly. Headless runs and clients using ACP no longer start swarms.
Moonshot opened the Pro and Highspeed variants in Kimi, Work, Code and its API. Its announcement scheduled the model weights for November 4.
Monitor and plugin monitors now work on Bedrock, Vertex and Foundry, including when nonessential traffic is disabled.
After retrying an occupied session, users can see which client holds it and choose whether to take it over.
Ten preset slots hold prompts or slash commands. Supported terminals can also display images in the transcript.
Admins can read or change Slack channel policies and attach access bundles through an organization API.
People required to migrate can export conversations and use their personal workspace during the grace period before joining the managed workspace.
Admins can review adoption under Usage and export a by-user report. Benchmarks and surveys move under Insights.
Opening a document in one window no longer changes another session’s view. Closing the last terminal tab now ends its background process.
This fix closes a gap that let shell initialization run outside the sandbox on macOS and Linux.
Anthropic lists general reliability work for this patch and gives no change-by-change description.
Teams can use Google Cloud IAM and Marketplace billing to reach Anthropic’s API, including fast mode on Opus 5.5.
Applications can send up to 400 questions in a request, reuse cached input and pass structured JSON or images.
Version 1.2026.286 adds controls above the composer for supported remote hosts. You can also mention chats on other connected hosts.
A new command helps migrate integrations to the October 21 API version, and safety decisions can appear in OpenTelemetry.
The geographic expansion opens additional markets for advertisers running campaigns in ChatGPT.
Plugins can use identically named skills without one hiding the other. Changing models mid-session now offers a choice about keeping or compacting context.
New major client SDKs and the ant CLI default to the October 21 API version. Existing integrations have a migration guide for breaking changes.
Owners and administrators can manage members and available controls from the shared console.
Pro and Max subscribers can try a separate Claude app that works proactively across ongoing tasks.
MoMo is available for Go, Plus and Pro purchases made on the web in Vietnam.
Organizations can limit the available service tiers, and individual projects can apply their own settings within those limits.
Platform operators can query whether an end customer has an open restriction or ban.
This patch fixes cloud sessions incorrectly reporting a missing login and avoids misleading prompt-length errors with large MCP toolsets.
The CLI enforces read-only settings in untrusted workspaces and fixes failures that could duplicate or erase resumed session history.
An optional sandbox setting supports programs such as Chromium and rootless Podman that create nested user namespaces.
Admins can disable web fetch in Console. Managed Agents also enforce organization-level fetch domain rules, including for redirects.
Enterprise customers can now use inference hooks outside the beta program.
A gradual rollout lets users hide a conversation from ordinary history and search, then reopen it with a PIN or supported device authentication.
Admins can register a webhook to receive notifications about end-customer safety enforcement.
User profiles let a platform associate Claude requests with a customer so an enforcement action can apply at the customer level.
Users can browse plugins and agents can look up tools as needed. Remote processes can keep running when a local client disconnects.
When an earlier plugin skips a hook, the transcript names both the plugin and the event.
Anthropic published a maintenance release without itemizing individual fixes.
Applications can search the web and fetch source pages through the same search engine used in DeepSeek’s products.
The deepseek-flash alias now serves the 1017 revision. Existing integrations retain the same calling convention and prices.
An environment setting selects the model used for client-side auto-mode classification.
OpenAI points workspaces to real-time Drive, SharePoint and Teams apps. Supported admins gain controls over content available through selected connections.
Start a presentation from files, a conversation or an outline, then continue editing it with collaborators in Space.
API customers can create a voice for GPT-Live and use nine additional voices across five languages.
Eligible users can watch parallel agents work and inspect their completed tasks. The update also improves response speed and tool use.
OpenAI sandboxes include tools for spreadsheets, slide decks and Word documents. Supported agents can also request faster output generation.
Global admins can set organization defaults, inspect connected users and revoke an app’s access in Admin Console.
Send a follow-up while Codex is working. A new settings browser lets you search preferences and see which changes require a restart.
Operators can set how long retry mode will wait for a usage limit to reset before ending a request.
Eligible labs can request sponsored ChatGPT Business Standard seats. Rosalind model access requires separate trusted-access approval.
Administrators in supported Business, Enterprise and Edu workspaces can change several plugin settings and review each result.
Enterprise compliance tools can retrieve instructions employees saved in Claude and Cowork.
Eligible admins can govern creation and sharing, connect private services via tunnels, and use US or EU regional hosting for new Sites.
The voice models add more natural accents across languages. New voice choices vary by account and region.
A capability field lets clients check whether a model can reason between tool calls before setting that mode.
Fast mode no longer requires access approval or a beta header for Opus 5.5 and is also available on supported Azure-hosted Foundry deployments.
Subscribers can create more Sites within their remaining usage and storage allowances.
GPT-Live can delegate decisions directly or let applications trigger them as incoming or outgoing transcript fragments arrive.
API request traces now separate five-minute and one-hour cache writes, making cache behavior easier to inspect.
Users can opt into progress messages and use the new kimi app command to start a desktop chat in a project directory.
Teams can follow Messages API requests and Managed Agents sessions in the observability system they already use.
The retirement applies to ChatGPT sign-in. Developers using the OpenAI API retain access.
Speech-to-speech sessions using grok-voice-latest move to 3.0. The 2.0 version is deprecated, and xAI lists the same price for 3.0.
Search results must satisfy the organization’s domain policy as well as the agent’s own allowlist or blocklist. Anthropic dates this change to October 14.
The listed GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots are scheduled to leave the API on January 14, 2027.
Managed Agents running through Claude Platform on AWS can use scheduled deployments.
Claude Code can compact long inactive conversations before the prompt cache expires to reduce reprocessing on the next turn.
The agent serializes file operations, writes files atomically and respects the resolved folder trust state in headless sessions.
Users can approve a coding plan, return to chat and review the result after Antigravity builds and tests it in the background.
Developers can request gemini-4-argon for complex coding, math and science tasks. The same day, Google began rolling it out to Pro and Ultra subscribers in the Gemini app.
The Customize menu brings skills and app connections together, with more than 40 supported services at rollout.
Version 1.2026.279 opens file citations in the viewer, shows Markdown files in reading view and lets you collapse the Projects section. This build also supports opening your dot through iOS widgets and Shortcuts.
Z.ai published recovery steps for 3.15.1 users. The patch also fixes failed requests when agents call models in parallel.
xAI began rolling out one subscription with a common usage pool for Grok and Cursor.
Built on GLM-5.3-Flash, the model was trained on legal research and multi-step tasks involving case records and legal documents.
Operators can adjust concurrency during a run and save workflows for replay. Autonomous Mode now asks before running dangerous commands.
No dated update recorded from the tracked sources.
No dated update recorded from the tracked sources.
The wire follows these official release notes, repositories and announcements.
Checked automatically: Codex CLI, Claude Code, Gemini CLI, Kimi Code, ZCode, Codex changelog RSS, OpenAI API changelog, Claude Platform release notes, Claude app release notes, Gemini API release notes, Gemini Apps release updates, Z.ai model release notes, Kimi research announcements, DeepSeek API change log, xAI dated API changes.