One changelog for all your AI dev tools.

30 Oct

• GPT-6 Rosalind model released for life sciences research↗

  OpenAI released GPT-6 Rosalind, a specialized model for life sciences research with scientific tool integration, available through a trusted-access program for authorized users.

26 Oct

• GPT-6.1 Terra model available for visual reasoning and computer use↗

  OpenAI released GPT-6.1 Terra, optimized for visual reasoning, computer use, and natural writing. Call it via standard chat and responses endpoints.

30 Oct

• Fast mode deprecated for Claude Opus 4.8, removed November 30↗

  Anthropic is removing fast mode from Opus 4.8 on November 30. Requests using speed: "fast" will fail; switch to Opus 5.5 to keep the feature.

29 Oct

• Usage and Cost Analytics API replaces older usage reporting endpoints↗

  Anthropic launched a new Analytics API with richer filtering (model, workspace, API key, user/service account) to replace two older usage-reporting systems. Existing integrations continue to work but should migrate to the new endpoints.

28 Oct

• Messages API and Token Counting API request size limit raised to 128 MB↗

  The Messages API and Token Counting API now accept requests up to 128 MB, up from 32 MB, letting you send larger documents and context in a single call.

31 Oct

• Disable ultrathink mode with an environment variable↗

  You can now disable ultrathink reasoning globally via `CLAUDE_CODE_DISABLE_ULTRATHINK` instead of managing it per-session.

• Fix shell commands left running when Claude Code exits↗

  Fixed shell commands remaining running on macOS and Linux when Claude Code exits unexpectedly, such as after credential refusal or cloud session archival.

• Set default effort level for sub-agents independently↗

  Sub-agents now have their own configurable default effort level via `CLAUDE_CODE_SUBAGENT_EFFORT`, decoupled from the parent session's setting.

• Fast mode for Opus 4.8 is deprecated and will be removed↗

  Fast mode for Opus 4.8 is deprecated and will be removed November 30, 2026. After removal, fast-mode requests will error instead of falling back. Switch to Opus 5.5 to keep the speed benefit.

30 Oct

• Idle conversations are now summarized to preserve context when cache expires↗

  If your conversation sits idle and the prompt cache expires, Claude Code now summarizes it automatically so the next message still has context, instead of losing history.

• Tool search requests are retried without tool blocks when gateway rejects them↗

  When a gateway repeatedly refuses requests because it doesn't support Claude Code's tool search format, the request is now sent once more without tool search, and the session remembers this limitation.

• Effort level is automatically lowered when admin caps exceed available capacity↗

  When your organization's admin caps a model's effort level below what Claude Code requests, it now automatically retries with a lower level instead of failing repeatedly.

• Bedrock sessions fall back to older model variants when default is unavailable↗

  If your Bedrock resource doesn't have the latest Sonnet, Opus, or Haiku model deployed, Claude Code now falls back to an older version of that model family instead of failing.

27 Oct

• Claude now reads entire large files before editing them↗

  Claude Code now detects when it has only seen the beginning of a large file and asks permission to read it completely or edit in place, preventing accidental overwrites.

• Skills now work after changing directories in a git repository↗

  Fixed a bug where skills scoped to specific directories or paths would stop working after you changed into a subdirectory of a git repo.

• Click to send queued messages in fullscreen mode↗

  In fullscreen mode, you can now click the hint text under a queued message to send it immediately, instead of only using the keyboard shortcut.

• Effort level changes mid-task now show correctly in UI↗

  Fixed the UI showing the new effort level you just selected before Claude actually switches to it; it now displays the level currently in use.

26 Oct

• Write and Edit operations no longer stall after large file changes↗

  Fixed Write and Edit stalling for up to 5 seconds after modifying thousands of lines in a file and incorrectly reporting zero lines changed.

• Interrupting Claude now takes effect immediately instead of waiting for timeout↗

  When you interrupt Claude while an MCP server is fetching a resource or a skill is downloading, the interruption now takes effect immediately rather than waiting for a timeout to expire.

• Conversation continues automatically when requests exceed token limits↗

  When a request exceeds 200K tokens, Claude Code now automatically compacts the conversation and continues instead of halting with a usage-credits error.

• Subagents can now use MCP tools from servers that connect after startup↗

  Subagents can now use MCP tools from servers that complete their connection after the subagent has already started, fixing a timing issue that previously blocked access.

30 Oct

• Sandbox session gains file operations and consistent path resolution↗

  Sandbox sessions can now inspect and manipulate files directly. Relative paths consistently resolve against the working directory across all methods, and just-bash now rejects directory overwrites.

• Partial JSON ending with minus sign in arrays now parses correctly↗

  Streaming JSON parsing now handles edge case where a minus sign (start of a negative number) appears at a chunk boundary in an array.

29 Oct

• Next.js Edge Runtime no longer warns on telemetry tracing↗

  Telemetry tracing no longer triggers false warnings when running on Next.js Edge Runtime.

• filterActiveTools graduates from experimental to stable API↗

  filterActiveTools moves out of experimental status with a deprecated alias for backward compatibility.

• LanguageModelStreamPart type graduates to stable↗

  The stream-part type definition is now stable with a deprecated experimental alias for backward compatibility.

27 Oct

• Remove runtime validation of inputs TypeScript already guarantees↗

  Runtime validation of message shapes and call settings is removed since TypeScript types already enforce them; value constraints types cannot express (ranges, non-empty collections, cycles) still validated.

26 Oct

• Add error and abort callbacks to text generation and agent loops↗

  New onStepError callback fires on individual step failures in generateText, streamText, and ToolLoopAgent; onError and onAbort callbacks added to ToolLoopAgent for full lifecycle error handling.

• Include SDK version in telemetry events↗

  Telemetry events now include aiVersion so tracing subscribers can identify which SDK version emitted each event, even in bundled or multi-version deployments.

• Add experimental steering to text generation and agent loops↗

  Experimental steering feature now available in generateText, streamText, and ToolLoopAgent for early testing.

30 Oct

• Geist Serif display typeface joins the font family↗

  Geist Serif is a high-contrast display serif built to the same cap and x-heights as Geist Sans, so you can mix them in the same line without moving baselines. It includes a variable axis for four dot-shape variants and 894 glyphs covering 352 languages.

• Data Cache graduates to Runtime Cache with per-project storage allocation↗

  Data Cache is being replaced by Runtime Cache with better isolation—each Pro project gets its own storage pool instead of sharing team-wide. Storage increases 8–10x depending on plan. Reads and writes become billable after free monthly allowances (1M reads, 200k writes on Hobby; Pro draws from $20 infrastructure credit). Automatic migration starts late November.

• Runtime Cache adds deployment markers and revalidation visibility↗

  New observability tools let you see why cache entries refresh, compare cache behavior to releases, and spot data written but rarely read.

• Cache reads use circuit breakers to avoid latency during traffic spikes↗

  During traffic bursts, checking whether cached data is still fresh could block reads. Circuit breakers now let reads proceed with briefly-stale freshness info, cutting p99 latency to one-fifth at comparable load.

• Data Cache merges into Runtime Cache with unified storage budgets↗

  Vercel consolidated its two caching systems into one. Next.js apps automatically migrate; storage budgets increase significantly per project, reducing evictions.

• Ling 3.1 Flash 1019 model available on AI Gateway↗

  InclusionAI's Ling 3.1 Flash 1019 is now accessible through Vercel's AI Gateway with a 1M-token context window, optimized for coding and tool-using agents. Free access ends November 3.

• Daemons can call services through declared bindings↗

  Daemons can now invoke services in the same deployment via bindings that inject the service URL as an environment variable. Vercel handles routing and TLS; bindings are one-way and survive promotions and rollbacks.

• Daemons now support HTTP liveness checks for automatic restart↗

  Configure an HTTP endpoint in your daemon to report health status. Vercel periodically polls it and restarts the container if the check fails, catching hung processes still running but unresponsive.

29 Oct

• Strix penetration testing tool integrates with Vercel AI Gateway↗

  Strix, an open-source AI penetration testing tool, now supports Vercel AI Gateway, letting you test against any of 400+ models from multiple providers through one API without separate provider accounts.

• IP geolocation resolves to city level and postal code for most traffic↗

  Geolocation data now resolves to city and postal-code granularity for most visitors, refreshed regularly to reflect IP range changes. Existing code works unchanged, but some VPN and hosting-provider traffic may resolve to different locations.

• Vercel Plugin available in Antigravity marketplace↗

  Antigravity agents can now discover and use the Vercel Plugin from the marketplace, gaining skills for Vercel Functions, Workflows, and other Vercel products to guide deployments.

28 Oct

• Box Mount integrates with Vercel Sandbox for agent file operations↗

  Agents running in Vercel Sandbox can mount Box folders as directories and sync changes back after the sandbox stops. The sync process runs as a separate user, isolating Box credentials from agent code.

• Run Paseo agents on Vercel Sandbox↗

  Paseo agents can execute in Vercel Sandbox instead of locally, letting you control sessions remotely and persist state across restarts without re-entering credentials.

• Agents can enable Web Analytics and Observability Plus non-interactively↗

  Agents can now programmatically enable Web Analytics and Observability Plus on Vercel projects without requiring interactive terminal sessions, enabling fully headless automation.

• fx coding agent now available as T3 Code provider↗

  You can now use fx, an open-source coding agent, within T3 Code's control plane on web, desktop, and mobile by adding it as a provider in Settings.

27 Oct

• fx adapter runs harness on host instead of in sandbox↗

  The fx adapter moved harness execution from the sandbox to your application's host process, eliminating the need for a sandbox bridge and port exposure. Model credentials remain on the host, improving security and reducing setup complexity.

• HarnessAgent accepts image inputs across all adapters↗

  Agents can now receive images alongside text through HarnessAgent's unified SDK, with vision processing delegated to each adapter's underlying model.

26 Oct

• Browse and search 1000+ services in Vercel Connect from the CLI↗

  Agents and developers can now discover and filter Vercel Connect services directly in the terminal without leaving the CLI, supporting programmatic discovery via JSON output for automation.

• Run Scale AI's AgentEnv evals on Vercel Sandbox↗

  Scale AI's AgentEnv testing framework can now run evaluations on Vercel Sandbox instead of locally, separating agent and environment execution for better isolation and reproducibility.

• Sandbox commands can now receive input through stdin↗

  Agent sandboxes can now send input to running commands through stdin, enabling interaction with stdio-based MCP servers and other CLI tools that expect live input.

28 Oct

• AI agents can now automatically optimize JavaScript bundles↗

  Next.js 16.4 adds an agent skill that analyzes your app's module graph to find unnecessary JavaScript—unused code, duplicate dependencies, missed tree-shaking, and deferrable components—then proposes or applies fixes and opens pull requests.

• Bundle analyzer shows critical render path and compares snapshots↗

  The interactive bundle analyzer gained a route summary homepage, sortable table view by size, critical render path highlighting, and automatic snapshot diffing to measure optimization impact.