One changelog for all your AI dev tools.

10 Sept

• Set expiration dates on project API keys↗

  You can set expiration dates when creating project API keys, and administrators can enforce maximum key lifetimes in Platform settings to require rotation.

8 Sept

• Prompt Cache Diagnostics now available in Responses API↗

  Responses API now includes diagnostics to measure cache reuse and identify why cache misses occur, letting you optimize prompt caching behavior.

10 Sept

• Claude Managed Agents permission policies add automatic evaluation mode↗

  Managed Agents can now auto-evaluate tool calls server-side instead of requiring explicit allow/deny policies. The server decides whether to run, block, or pause each call; events report the evaluation decision.

• CLI can attach to and control Managed Agents sessions in real time↗

  The `ant` CLI now lets you connect to a running Managed Agents session from your terminal, watch it execute, send new messages, and manually approve or deny tool calls waiting for approval. `--web` opens the session in a local web viewer.

11 Sept

• Concurrent agent limit in Workflow tool is now configurable↗

  Set `CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS` to tune how many agents the Workflow tool can run in parallel, up to 256, for better throughput on inference-heavy tasks.

• Bash tool shows diffs of files it edits↗

  When the Bash tool edits files, it now shows you a diff of what changed, making it easier to review file modifications without re-reading the whole file.

• Output styles can be switched in headless and remote sessions↗

  You can now change output styles mid-session in headless and remote deployments using `/output-style`, matching the interactive terminal experience.

• Plugin eval suite scores plugins against Claude Code reproducibly↗

  Plugin authors can now run `claude plugin eval` to test their plugins against Claude Code with reproducible, scored results in JSON and HTML formats.

10 Sept

• Third-party Anthropic-compatible endpoints now work with artifacts↗

  Artifact tool was rejecting requests to third-party Anthropic-compatible endpoints due to a schema regex; now works with any compatible API.

• Stale model-access cache no longer silently switches models mid-session↗

  If another Claude Code process refreshed the model-access cache while you were in a session, your session could silently switch to the organization's default model. Now stays on your chosen model.

• Symlinked directory permissions now apply consistently↗

  Permission rules on symlinked directories (like `/etc` or `/tmp`) were bypassed if the real path was used instead of the symlink. Now enforced consistently.

• WebFetch now times out instead of hanging indefinitely↗

  WebFetch could hang forever on servers that don't close connections. Now times out after 300 seconds; set `CLAUDE_CODE_WEBFETCH_DEADLINE_MS` to change the deadline.

9 Sept

• Restore prompt-cache stability across session resume and model switches↗

  System prompt and tool definitions are now cached once per session instead of re-rendering on resume or model change, reducing latency and token waste for long conversations.

• Render system prompt fresh on each request↗

  Use `--system-prompt-snapshot off` to iterate on prompt text without the agent caching and replaying the old version from earlier in the conversation.

• Cap effort level across all model providers↗

  Add `maxEffortLevel` to your config to enforce a ceiling on reasoning effort across Bedrock, Vertex, Foundry, and other providers, while letting users dial down per request.

• Fix Cowork scheduled tasks failing with required sandboxing↗

  Organizations with mandatory sandboxing policies can now run Cowork scheduled tasks without startup failures.

8 Sept

• Gateway environment variable no longer forces Cloud sign-in when API key is set↗

  A 2.1.265 regression made CLAUDE_CODE_USE_GATEWAY force Cloud-gateway authentication even when an API key was configured, breaking proxy setups. Now it requires both base URL and token again.

• Resume after process death keeps interrupted tool calls and avoids prompt rewrite↗

  If Claude Code crashes mid-tool-call, resuming now preserves the interrupted call and the prompt context, avoiding unnecessary rewrites.

• Plugin directories now load all child folders with manifests dynamically↗

  You can point --plugin-dir at a folder and each subfolder with a manifest loads as a plugin. Changes to the directory are detected while the session runs.

• Subagent resumption now preserves tool list and system prompt for cache reuse↗

  When you resume a subagent, its tools and system prompt now stay the same, which lets Claude Code reuse the cached prompt context instead of recomputing it.

• Tool results are capped at 1 GB and truncation is shown in preview↗

  Large tool outputs are now limited to 1 GB on disk. The preview in the conversation shows you when a result was truncated so you know it's incomplete.

11 Sept

• Batch requests now support image generation↗

  You can now include image generation requests in batch API calls, expanding what workloads batching supports alongside text and embedding operations.

• Invalid reranking provider indices are now rejected↗

  The SDK now validates reranking provider indices and rejects requests with invalid values, preventing silent failures.

• Embed and rerank operations now track runtime context↗

  Embedding and reranking operations now capture and attribute runtime context, improving visibility into where these operations are called from in your application.

• Tool output errors now have a stable type and guard↗

  A new stable type and type guard are available for tool output errors in UI messages, making error handling more reliable and type-safe.

9 Sept

• Batch cancel and list APIs for managing multiple requests↗

  New batch cancel and list APIs let you manage groups of concurrent requests—useful for controlling cost and latency in high-volume inference.

• Batch cancel and list APIs for managing concurrent model calls↗

  New batch cancel and list APIs let you manage groups of concurrent model calls, inspect their state, and cancel them in bulk.

• Embedding models now reject responses with no embeddings↗

  Embedding model responses that contain no embeddings are now rejected, preventing silent failures and making errors explicit.

8 Sept

• Tool choices are now enforced in streamText↗

  The streamText function now respects tool choice constraints during streaming, preventing the model from calling tools outside the allowed set.

• Batch API now accepts per-request model selection↗

  The batch API now lets you specify a different model for each request in the batch, instead of requiring all requests to use the same model.

11 Sept

• Vercel Sandbox storage doubles to 64 GB↗

  Vercel Sandbox now provides 64 GB of storage instead of 32 GB, giving more room for large builds, dependencies, and disk-heavy workloads.

• Agents can run in isolated sandboxes with tailnet identity and zero data retention↗

  Agents now execute in Vercel Sandbox with identity validated through the tailnet, no API keys issued to the agent itself, and optional zero-data-retention enforcement per request. This isolates agents that can read private data and act on it from the public internet.

• Eve agent framework provides durable sessions and streaming out of the box↗

  Eve's useEveAgent hook eliminates boilerplate for agent implementation, bundling durable sessions, streaming, tool calls, and approval prompts with automatic model routing through AI Gateway.

• Workflow SDK enables long-running background jobs without separate orchestration↗

  Vercel's Workflow SDK lets you run long-running, multi-step background jobs directly on the platform instead of managing separate orchestration infrastructure, reducing operational overhead.

• AI Gateway now powers model routing through network identity↗

  Tailscale built Aperture, a model router that grants and revokes model access through tailnet identity instead of API keys. AI Gateway provides the unified routing layer across hundreds of models, with cost and usage returned per request so Aperture can surface spend without maintaining provider price tables.

• AI Gateway and AI SDK unify model routing across multiple providers↗

  AI Gateway and AI SDK let you route requests across multiple AI models through one standardized interface, so switching providers or testing new models becomes a configuration change instead of rewriting code.

• Control who can create and manage connectors in Vercel Connect↗

  Owners can now limit who creates and manages connectors in Vercel Connect, restricting access to Owner role or users with explicit extended permission. Useful for teams that need credential governance.

10 Sept

• GitHub Copilot available through AI SDK harness layer↗

  GitHub Copilot joins the AI SDK harness layer, which abstracts different coding agents behind a single interface. Switch agents by configuration instead of rewriting application code.

• FastAPI static files and frontends now served from CDN at build time↗

  FastAPI static assets and frontends are now built into the CDN instead of served through Functions, reducing compute costs and latency. Middleware and auth-guarded routes stay on the Function by default; override with tool.vercel.fastapi.static.cdn or .exclude.

• Deploy OpenAI agents with Vercel Sandbox for code execution↗

  You can now build agents on OpenAI's API that execute code and access files via Vercel Sandbox, with reliable reconnection through signed webhooks and persistent workspaces that scale to zero.

• Vercel Sandbox now runs in all 20 compute regions↗

  Sandboxes now deploy across all 20 Vercel regions for lower latency and data residency compliance. Pro and Enterprise teams can configure failover regions; existing sandboxes stay put.

• CDN metadata lookups now 91% faster through sharded caching↗

  The CDN now batches metadata for many paths into indexed shards instead of fetching each path's metadata separately. This cuts cache misses on redeployment and reduces lookup latency by 91% for large sites, plus speeds up the build step.

9 Sept

• Deployment Protection now free on all plans, including production↗

  Production deployments can now be protected behind Vercel Authentication for free on every plan. Previously this required the Advanced Deployment Protection add-on. Useful for internal tools and pre-launch sites.

• Eve agents can retain context across sessions with persistent memory↗

  Eve agents can now maintain context across conversations using persistent memory slots. Each slot specifies a storage provider and scope—for example, separate memory per authenticated user. Memory is automatically retrieved before each turn and can be updated after turns or via agent tools. Built-in file provider persists to Vercel Blob; custom providers are supported.

• CLI can read and search Vercel changelog feed↗

  Agents and developers can now query Vercel's changelog directly from the CLI to discover features and find updates relevant to their projects, with JSON output for scripting.

• DeepSeek V4.1 Flash available on AI Gateway with vision support↗

  DeepSeek V4.1 Flash is now available through Vercel's AI Gateway with vision capabilities, letting you send text and images together. It has a 1 million token context window and supports up to 384,000 token responses, plus reasoning, tool use, and prompt caching.

• v0 can connect to email, auth, search, and database providers inline↗

  When you prompt v0 to build something that needs a provider like Resend, Clerk, or MongoDB Atlas, it can connect directly in chat, auto-configure credentials and environment variables, and generate code using the provider's recommended patterns.

8 Sept

• Deployment step is now 10% faster↗

  Vercel now uploads one routing manifest instead of separate metadata files per function, saving ~1 second per deploy and up to 12 seconds for large apps.

• Vercel Sandbox domain routing is 18x faster globally↗

  Sandbox domains now resolve from the nearest regional replica, cutting median lookup latency from 62ms to 3.4ms. Regions far from the previous store see much larger gains—up to 112x faster in Sydney, 146x faster in Cape Town at p99.

• Flat Rate CDN pricing replaces usage-based billing for Pro teams↗

  Vercel Pro now offers predictable monthly CDN pricing with built-in spike protection. Teams choose a capacity tier matching typical usage; temporary traffic bursts don't trigger surprise bills or degrade service.