One changelog for all your AI dev tools.

5 Nov

• GPT-6-Cyber model available for security testing↗

  OpenAI released GPT-6-Cyber, a model designed to help investigate vulnerabilities, reproduce them safely in test environments, and verify fixes. Requires Daybreak setup before use.

4 Nov

• Markup-guided editing added for GPT-Image models↗

  Image editing now accepts markup instructions for GPT-Image 2.5 and 2.6, letting you specify edits more precisely than text alone.

• GPT-Image-2.6 models support SVG and animated GIF generation↗

  Two new GPT-Image-2.6 variants (Sunburst, Flare) generate SVGs and animated GIFs with tunable complexity, expanding beyond raster outputs.

2 Nov

• GPT-Live can now initiate outbound calls via SIP↗

  The Live API gained outbound calling capability through SIP integration, letting you build voice agents that initiate calls rather than only receive them.

5 Nov

• IP allowlist for API requests now available in beta↗

  Organizations can now restrict API requests to whitelisted IP addresses in the Claude Console (beta). This limits exposure if an API key is compromised by refusing requests from unexpected locations.

3 Nov

• Claude models now carry Anthropic's text watermark on supported platforms↗

  Claude's latest models now embed Anthropic's watermark in generated text on supported platforms, helping identify AI-generated content. On Sonnet 4.6, non-default sampling parameters may prevent watermarking.

2 Nov

• Message threads reduce payload size and latency in long conversations↗

  Message threads keep conversation history on Anthropic's servers, so you send only new messages each turn instead of the full history. Reduces latency and bandwidth for agents and long-running loops. Beta feature across Claude API, AWS, Google Cloud, and Microsoft Foundry.

6 Nov

• Auto mode now requires server-side review when telemetry is off↗

  If you've disabled telemetry or feature flags, auto mode will now block actions instead of falling back to a built-in classifier. Set the environment variable to restore the old behavior.

• Permission prompts can now be denied automatically instead of blocking↗

  You can now set a permission rule to automatically deny actions instead of prompting you. The turn continues without interruption. This works in terminal sessions, subagents, and teammate mode.

• Subagents now show finished tasks in a session history row↗

  The `/tasks` command now keeps finished subagents visible in an "Earlier in this session" section, so you can reopen one and continue messaging it even after restart.

• MCP tool calls can now run as background tasks↗

  When an MCP server runs a tool call as a task, Claude Code now backgrounds it right away instead of blocking. This lets Claude continue working while the task finishes. You can disable this with an environment variable if you need the old blocking behavior.

4 Nov

• Add effort level control to mod forks↗

  A mod can now spawn a quick side task at low effort even if the main session uses higher effort, letting you balance speed and cost per task.

• Show stored sign-in expiration in auth status and JSON output↗

  You can now check when your stored Claude sign-in will expire via `claude auth status` or programmatically in JSON output, and get a stderr warning 10 days before it does.

• Fix --continue losing conversation after malformed tool call retries↗

  A bug introduced in 2.1.314 caused `--continue` to lose prior messages when a malformed tool call ran out of retries; now fixed.

• Record user consent when Claude Code enables mods in a session↗

  When Claude Code asks permission to enable a mod during a session, your response is now recorded for compliance and debugging purposes.

3 Nov

• Prompt before sending account email to external services↗

  Web and shell tool calls now prompt for confirmation before sending your account email address to external services.

• Show when organization settings fail to load in Code tab↗

  Claude Desktop now shows you when your organization's remote settings are still loading or have finished loading in the Code tab, so you know the state.

• Warn when scheduled tasks gain unexpected connectors↗

  When you schedule a cloud task, Claude now warns you if it adds connectors you didn't explicitly ask for and shows how to remove them.

• Block actions when Agent SDK hook callbacks fail↗

  Agent SDK hook callbacks can now block the action they're attached to if the callback itself fails, instead of silently letting the action proceed.

7 Nov

• Stop leaking provider credentials in telemetry and traces↗

  Provider URLs containing credentials are now stripped before being sent to telemetry and tracing systems, preventing accidental exposure of authentication tokens.

6 Nov

• SDK respects toolChoice: 'none' even when tools are empty↗

  When you set toolChoice: 'none', the SDK now preserves it even if you don't pass any tools to the SDK itself. This lets provider-level tool configurations be properly disabled instead of overridden.

5 Nov

• Audio type detection from bytes when servers return generic content types↗

  The SDK now infers audio types from file bytes when servers don't provide proper content-type headers, fixing transcription with generic responses.

• Union type inference for structured output schemas↗

  Passing a union of schemas to structured output now correctly infers the output type as a union, instead of only using the first schema's type.

• Native URL transcription without downloading for supported providers↗

  Providers that support native URL transcription (like Groq) can now process audio URLs directly without downloading them first, avoiding upload limits. Telemetry events report `undefined` for `byteLength` and `mediaType` when providers fetch URLs natively.

• Typed streaming heartbeat parts with Anthropic support↗

  Heartbeat pings from Anthropic are now typed stream parts instead of raw chunks, and consumers with exhaustive pattern matches must add a `ping` case when upgrading.

3 Nov

• Tool metadata restored in approval continuations↗

  Approval policies now receive the same tool definition fields during continuation validation as they do for new calls, fixing inconsistency in tool context.

• Durable workflow telemetry with OpenTelemetry support↗

  Durable WorkflowAgent runs now emit one complete telemetry event on completion or suspension, with original timestamps preserved. OpenTelemetry integration requires explicit bridge setup to carry trace context across steps and nest model spans under provider requests.

7 Nov

• DeepSeek V4.1 Flash 1017 Ultrafast model available on AI Gateway↗

  DeepSeek's October 17 Flash update is now accessible through Vercel's AI Gateway with 1M-token context and 1k tokens/sec throughput, optimized for interactive coding and agentic workflows.

6 Nov

• Agents can stream generative UI components via JSON specs↗

  Agents can now generate UI by writing JSON specs from a typed component catalog, which Chat SDK converts to native cards for each platform. Validation catches invalid specs before sending; platform limits can be enforced or degraded gracefully.

• AI Gateway adds DeepSeek Search and Fetch tools for web retrieval↗

  AI Gateway now includes DeepSeek Search (finds pages with summaries) and Fetch (retrieves Markdown content), usable with any model supporting tool calling. Switch providers without changing tools.

• Vercel CLI can set projects to auto-delete after a duration↗

  New --delete-project-after flag on vercel deploy auto-deletes projects after a set duration, useful for ephemeral prototypes and test deployments that don't need manual cleanup.

• Vercel Tracing moves to general availability with automatic setup for new projects↗

  Vercel Tracing is now GA. New Next.js projects with Node.js functions automatically collect traces (100% sampling) with no setup required; you can search across your team's projects via dashboard or CLI, and trace retention increased (Hobby: 3 days, Pro/Enterprise: 14 days).

• Trace search now covers your entire team across all projects↗

  The new Traces page shows every trace and span across your team or a single project, searchable with KQL. Live mode streams new traces in real time. Same queries work in the CLI with vercel traces search.

• Token Pass credit now applies automatically across AI Gateway↗

  Token Pass subscribers can now use standard API keys or OIDC tokens to access included model credits automatically, with fallback to team credits when exhausted. Dedicated Token Pass keys remain optional and stop at period end.

• Ampersand model available on AI Gateway through mid-November↗

  Humans& released Ampersand, a model trained for team chat that decides when to speak or stay silent. Available free on Vercel's AI Gateway until November 13; use it in your app or coding agents via standard APIs.

5 Nov

• Vercel Schedules moves to public beta across all plans↗

  Schedules is now generally available in beta on all plans, replacing Cron Jobs. Define static schedules in config or create them at runtime; both can be managed without redeployment. Existing Cron Jobs continue working.

• Gemini Omni 1.2 Flash video generation available on AI Gateway↗

  Gemini Omini 1.2 Flash is now available through Vercel's AI Gateway, letting you generate and edit short videos via a single API without a separate Google account. Supports up to 4K output and video extension up to 30 seconds total.

4 Nov

• Workflow SDK makes idempotent starts atomic with hook claiming↗

  Workflow start() is now atomic: it admits the run and claims the hook in one step, closing a documented race condition where duplicates could slip through the gap between lookup and start. Use your domain's ID (order, invoice, request) as the idempotency token.

• Kimi K3.1 Pro available on Vercel AI Gateway with Zero Data Retention↗

  Moonshot's Kimi K3.1 Pro models are now accessible via Vercel's AI Gateway with Zero Data Retention support, letting you enforce data residency policies per-request or across your team.

• Agents can create and cancel schedules without redeploying↗

  Agents can now create schedules at runtime instead of requiring static cron config, enabling user-driven scheduling like 'run this tomorrow' without database polling.

• Private functions can run scheduled tasks without HTTP exposure↗

  Scheduled tasks can now call private functions that are not exposed as HTTP endpoints, with built-in verification that requests come from the scheduling service.

• Vercel tokens are now automatically scanned and revoked when exposed publicly↗

  Vercel now automatically scans public sources like GitHub, npm, and Docker for exposed tokens and revokes them before they can be misused. You'll get an email if your token is revoked, preventing unauthorized access and unexpected costs.

• Test scheduled tasks locally and on preview deployments↗

  You can now test scheduled work locally during development and on preview deployments, with a Run Schedule button to trigger tasks immediately without waiting.

3 Nov

• Diagnose Vercel CLI setup problems with vercel doctor↗

  Vercel CLI now includes `vercel doctor`, which diagnoses setup issues and suggests fixes for each problem. It works before login or project linking, making it safe for agents to run first. Use `--json` for structured output.

• Preview deployments and rollback for each branch enable rapid iteration↗

  Each branch gets its own preview URL for testing before production merge. The team ships multiple times daily and can instantly revert if needed, critical for a medical application.

2 Nov

• Findings are available via CLI for integration with your own coding agents↗

  Findings can be retrieved via CLI as JSON or Markdown, letting you pipe them into your own coding agent for fixes. Dashboard findings also copy as text for pasting into agent chats.

• LongCat 2.5 model now available on AI Gateway↗

  Meituan's LongCat 2.5 is now accessible through Vercel's AI Gateway, supporting text and image inputs with extended context for code-heavy agent tasks across complex projects.

3 Nov

• Binary arguments to `use cache` can produce identical cache keys↗

  Apps that use binary arguments (buffers, typed arrays) in `use cache` to distinguish users or accounts risk cache collisions where different inputs produce the same key, causing cross-user data leaks or poisoning.

• Nested `use cache` functions leak content across routing boundaries↗

  Apps with nested `use cache` functions that read root parameters (tenant, site, locale) can serve one value's content to requests for another. The outer cache entry stores before the inner function's dependencies are fully tracked. Requires immediate patching.

• Cache isolation broken between tenants in concurrent `use cache` calls↗

  Multi-tenant apps using Cache Components risk serving one tenant's cached data to another when `unstable_cache` and `use cache` are combined and read root parameters like tenant IDs. Upgrade to v16.4.5 or v15.5.29.

• Cookie values exposed in internal response header on Edge routes↗

  Edge routes setting cookies via `NextResponse.cookies.set()` expose the cookie value in an internal header readable by JavaScript, bypassing `HttpOnly` protection. Requires attacker-controlled script on the same origin. Node.js routes and `cookies().set()` are unaffected.