Your stack
Models
2 pickedAI coding tools
1 pickedAI frameworks
1 pickedAuth & data
1 pickedInfra & queues
1 pickedCloud platforms
4Frontend & UI
1 pickedLanguages & runtimes
7Observability & evals
2Section order
drag to arrange- OpenAI1
- Anthropic2
- Claude Code3
- Vercel AI SDK4
- Vercel5
- Next.js6
- Drizzle ORM7
Verbosity
How often
Your edition
73 items this week
5 Oct – 11 Oct
OpenAI
3 updates
8 Oct
GPT-6.1 Sol adds ultrafast mode for lower token latency
GPT-6.1 Sol now supports ultrafast mode via the Responses API, reducing latency between generated tokens for all API users subject to rate limits.
7 Oct
chat-latest snapshot now tracks ChatGPT Plus, Pro, Business and Enterprise model
The chat-latest model alias now updates to track the newest model deployed in ChatGPT for paying tiers. OpenAI recommends GPT-6 for production but chat-latest lets you test latest improvements without pinning versions.
6 Oct
Decisions API in beta for structured output at higher speed
New Decisions API (beta) with gpt-6-luna model outputs structured data 10x faster than the Responses API, useful for high-throughput structured extraction.
Anthropic
8 updates
9 Oct
Claude Managed Agents can orchestrate dynamic workflows
Managed Agents can now write and execute multi-phase workflows for large-scale tasks like batch document review, with the server orchestrating agent runs and combining results.
8 Oct
Compliance API now returns chats from unified Claude experience
The Compliance API can now retrieve chats created in the unified Claude experience for Enterprise organizations in beta, expanding audit and data-retrieval scope without requiring new credentials.
7 Oct
SDK adds browser and computer use tool classes in beta
The Python and TypeScript SDKs now include beta classes for browser and computer use tools. You subclass one, implement one method per tool, and the SDK handles the tool loop, URL/file policies, and approval callbacks.
Managed Agents restrict web tools to allowed hosts in limited networking
Managed Agents with limited networking now apply allowed_hosts to web_search and web_fetch tools. Mismatched allowed_domains entries block session creation with a 400 error. web_fetch also now only fetches URLs that appeared earlier in the session.
Claude Haiku 5.5 launches with 1M context and adaptive thinking
Haiku 5.5 is the new fast model with 1M context window, adaptive thinking (on by default), and 128k output tokens. Code written for Haiku 4.5 breaks: manual extended thinking returns 400 errors and responses may start with thinking blocks.
Prompt cache reads on Sonnet 5.5 cost half as much
Cache reads on Sonnet 5.5 now cost $0.10 per million tokens instead of $0.20, cutting the rate in half. Cache writes and other prices unchanged.
6 Oct
Models API reports which models support web search and code execution
The Models API now reports whether each model accepts web search and code execution tools via `capabilities.server_tools`, letting you check model abilities before making requests.
5 Oct
Models API reports which models support disabling thinking
The Models API now reports whether each model supports disabling extended thinking via `capabilities.thinking.types.disabled`, letting you check before sending requests.
Claude Code
20 updates
9 Oct
Subagents can auto-compact their window independently of the main conversation
Subagents can now auto-compact their conversation window separately from the main agent, letting you tune memory usage per subagent.
Workflow agents can run on a single model while other subagents use theirs
Set `CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL` to run every workflow agent on a single model while other subagents keep their own model choices.
Read tool can access large text files when context allows
The Read tool now accepts an `allow_large` option to read text files past the usual size limit in a single call when context permits.
Managed hooks now correctly deny tool calls and block prompts
Fixed managed-settings hooks: `PreToolUse` hooks that deny with `"continue": false` and `prompt` hooks that block now correctly refuse the call while keeping the turn open.
8 Oct
Hook failures can now block actions instead of passing through
Hooks can now block actions on failure instead of letting them through. Set `onFailure: "block"` on command or HTTP hooks to stop execution if the hook fails to start, times out, or exits with a non-zero code.
MCP servers now back off on repeated connection drops
Remote MCP servers that drop connections repeatedly now back off exponentially instead of reconnecting in a tight loop, up to 30 seconds. Fixes headless and SDK sessions staying disconnected after outages.
Subagents now see their own git state, not the parent's
Subagents running in their own linked worktree now see their own git state instead of inheriting the parent session's branch, status and commits.
Prompt and agent hooks now properly enforce instruction-based blocks
Hooks written as natural-language instructions (e.g., "Block commands that...") were incorrectly permitting blocked actions; this is fixed.
7 Oct
Fixed Claude losing work after context compaction
Claude sometimes incorrectly marked its own actions as incomplete after the context window was compacted, causing it to redo or undo work it had already finished.
Tool schemas can now appear in prompts immediately instead of behind search
Mod authors can now set `isDeferred: false` when registering tools to include their schemas in the initial prompt instead of deferring them until tool search.
Subagent scripts can now distinguish custom agent types
Scripts that monitor subagents can now read the `agentType` field from status payloads to tell custom agent types apart.
6 Oct
Fixed sub-agents incorrectly entering auto mode when unavailable
Sub-agents now respect when auto mode is disabled by settings, circuit breaker, or model limitations, instead of incorrectly activating it anyway.
Plugin marketplace can be added during install
You can now add a marketplace and install a plugin from it in one command, with the same security checks as adding it separately.
Sub-agents now run at a specified effort level
You can now specify the effort level when spawning sub-agents, letting you trade speed for quality on specific tasks without changing global settings.
Prompt caching available for mod-written completions
Mods that call the completion API can now cache large prompt or system blocks, reducing latency and cost for repeated requests with the same context.
Session messages no longer lost when quitting
Quitting Claude Code was losing the final messages of your session in v2.1.288; now restored.
Cloud sessions now preserve answers to permission prompts
Cloud sessions were losing your responses to permission prompts in v2.1.290; now fixed.
5 Oct
Hook results now include server tool calls made by the advisor
The `turn.step` hook result now includes `serverToolUses`, exposing tool calls the advisor made during the turn so mods can react to them.
Subagent permission checks now include the organization's approval ceiling
The `tool.check` hook now includes `ceiling`, showing the organization's required approval tier for a tool, so mods can enforce policy.
Plugin validation now reports which hooks have error handlers
`claude plugin validate` now shows under `--json` which hooks have `.catch` error handlers, making it easier to audit plugin robustness.
Vercel AI SDK
18 updates
9 Oct
Resumed streams no longer carry over unrelated message state
When resuming a stream after interruption, the SDK was incorrectly preserving message state from previous unrelated calls. This is now discarded.
Streaming output timeouts no longer block on local tool execution
Model output timeouts (`chunkMs`, `firstChunkMs`) now pause during local tool execution, so slow tools won't trigger spurious timeouts. Each retry gets a fresh budget.
8 Oct
WebSocket chat transport with streaming and resume support
Add a WebSocket-based chat transport that maintains connection state, handles backpressure, and can resume interrupted streams using sequence numbers.
Top-level dimensions setting for embedding functions
Embedding functions now accept a top-level dimensions configuration, reducing boilerplate when all calls use the same embedding size.
Warning logged when tool loop hits step limit
The SDK will warn you when a tool loop reaches its step limit, helping you spot runaway loops during development.
Tool update failures in sendAutomaticallyWhen now propagate correctly
When a tool update fails asynchronously inside sendAutomaticallyWhen, the error now surfaces to your handler instead of being swallowed.
Provider-executed tool calls survive stream resumption
When you resume a stream after interruption, tool calls that the provider executed are no longer lost; they persist in the message history.
Decision state now supports ordered text, file, and JSON parts
Decision state now supports structured text, file, and JSON parts in a defined order. Raw arrays passed as state are treated as decision state parts; wrap JSON arrays in an object or json part to preserve them.
7 Oct
Add onUIMessageStepEnd callback to persist messages
onUIMessageStepEnd callback lets you hook into message lifecycle after UI steps complete, useful for saving conversation state.
Abort signal now cancels pending streams and rejects promises
When you abort a request, pending promises are rejected and the stream is cancelled, giving you proper cleanup instead of hanging operations.
Model refusals are now reported as refusal answers
When a model refuses a request, it's now reported as a refusal answer type rather than thrown as an error, letting you handle it as structured output.
6 Oct
Emit deprecation warnings with stable codes and deduplicate
Deprecation warnings now include stable identifiers and are deduplicated, making it easier to track and suppress specific warnings in your application.
Preserve V2 provider-reported total tokens in raw usage
Fixed regression where total token counts from V2 providers were lost; raw usage now retains provider-reported values.
Add portable max reasoning level with provider mappings
New `max` reasoning level works across providers with automatic mapping and fallback support. TypeScript consumers with exhaustive `ReasoningLevel` switches must handle `max` before upgrading; existing calls remain compatible.
5 Oct
Throw typed error for unrecognized audio formats
When generated audio has an unrecognizable media type, the SDK now throws a specific error type you can catch with AISDKError.isInstance().
Preserve tool approval state when resuming interrupted streams
When a stream pauses and resumes, the SDK now retains which tools were already approved, preventing re-approval prompts.
Support custom reasoning delimiters in extractReasoningMiddleware
You can now configure the markers that wrap reasoning text in extractReasoningMiddleware, instead of using only the built-in delimiters.
Add non-streaming WorkflowAgent.generate() method
You can now call WorkflowAgent.generate() for synchronous, fully-typed results instead of only streaming; it shares the same tool-calling loop and defaults to 20 steps.
Vercel
20 updates
9 Oct
Agents can search, price, and initiate domain purchases via CLI
Agents gain a structured workflow to discover and price domains, then hand off purchase decisions to you for approval. The CLI returns machine-readable responses so agents can chain commands without manual steps.
Pro teams default to 30-day deployment retention
Pro teams default to 30-day deployment retention to manage storage costs as agent-driven deployment frequency increases; you can extend it in settings.
eve framework keeps agent instructions, tools, and skills in one directory
eve is an open-source agent framework that lets non-engineers shape agent behavior through markdown instructions and skills, while tools live in TypeScript files whose names become the agent's API surface.
eve agents access Slack and Linear through reusable Vercel Connect tokens
Vercel Connect gives agents access to external services through short-lived tokens managed at runtime, eliminating the need to store provider credentials in each application and reducing the surface area for credential leaks.
Enterprise Managed Users and Directory Sync control agent builder access
Enterprise Managed Users lets teams manage Vercel access through their identity provider with SSO and Directory Sync, so account administration follows the company's directory instead of requiring manual Vercel account management.
AI Gateway routes model requests through one integration
AI Gateway lets teams route agent requests to multiple model providers through a single integration, with usage tracking and cost visibility per model.
Microsoft Decision-1 model available through AI Gateway
Microsoft's Decision-1 model is now available through Vercel's AI Gateway, letting you classify requests, route workflows, and score outputs with probability scores for each option in a single request.
Liquid AI's d1 decision model available on AI Gateway
Liquid d1, a decision model that classifies and scores without text generation, is now accessible through Vercel's AI Gateway. It supports vision input and returns structured answers with probabilities; calls integrate into your usage tracking and cost reporting.
8 Oct
Skip sending request bodies to Routing Middleware
Routing Middleware can now skip receiving client request bodies via skipMiddlewareRequestBody config, reducing bandwidth costs and latency when middleware doesn't need them; bodies still reach Functions and rewrite targets.
Sandbox.create() now waits for snapshot restoration to complete
Sandbox creation now includes snapshot restoration time, so the sandbox is ready to use immediately. Errors surface at creation time rather than on first operation, making error handling more predictable.
Step 5 Preview model available on AI Gateway
StepFun's Step 5 Preview model is now available through Vercel's AI Gateway, supporting 1M-token context and image inputs. Use it directly via API or route through coding agents like Claude Code and Cursor.
Grok Imagine Video 1.5 Lite model available on AI Gateway
xAI's lightweight video model is now accessible through Vercel's AI Gateway, supporting text-to-video and image-to-video workflows with native audio and multiple resolutions and formats.
7 Oct
Glyph Cluster reasoning model available on AI Gateway for coding tasks
Glyph Cluster is a new reasoning model on AI Gateway for coding and long-context work like debugging, planning, and synthesis. It supports function calling for agent tool use and streams responses, but lacks image input and structured output support.
OpenAI Decisions API available through AI Gateway
OpenAI's Decisions API is now available on AI Gateway, letting you route decision requests through the same gateway as other AI calls. Decision models return probabilities and scores instead of text, useful for routing and triage tasks.
Vercel Flags now support timestamp attributes for time-based targeting
Vercel Flags can now target users based on time windows—useful for campaigns, promotions, or feature rollouts tied to dates. Configure attributes in the dashboard or CLI, pass Date.now() from your app, and set rules with date pickers.
Claude Haiku 5.5 available on AI Gateway with adaptive thinking
Claude Haiku 5.5 now runs on Vercel's AI Gateway with adaptive thinking support. Set reasoning_effort to control how much the model thinks and token consumption, useful for cost-sensitive tasks like summaries and classification.
6 Oct
AI Gateway escalates to fallback models based on answer confidence
AI Gateway now lets you define confidence conditions that trigger fallback models when primary answers are uncertain. Useful for routing low-confidence Choice, Score, or Boolean decisions to a more capable model without changing existing fallback logic.
Mistral Large 4 available through AI Gateway
Mistral Large 4, a multimodal reasoning model, is now accessible through Vercel's AI Gateway without a separate Mistral account, using standard OpenAI-compatible and Anthropic APIs.
Nano Banana 2.1 image model available on AI Gateway
Google's Nano Banana 2.1 image model is now accessible through Vercel's AI Gateway with improved image generation and editing capabilities including scene recontextualization and region-specific mask-based edits.
5 Oct
OpenAI Ultrafast tier available on AI Gateway for GPT-6 Astra
AI Gateway now exposes OpenAI's Ultrafast service tier for GPT-6 Astra, enabling faster output for interactive apps. Ultrafast covers US and global regions; EU requests fall back to standard tier. Costs 6× per-token.
Next.js
3 updates
6 Oct
Cache Components now recommended for all Next.js apps
Cache Components, which let you mark parts of your component tree for caching via `'use cache'` directives, are now stable enough to recommend for every app. New projects use them by default; existing apps can migrate with `next upgrade --agent`.
ensureStatic prevents accidental dynamic content in routes
New `ensureStatic` export lets you enforce that a route, prefetch, or shell contains only static content, failing the build if a dynamic component sneaks in. Useful for cost-sensitive sites like ecommerce or blogs.
Exclude specific content from link prefetches
New `exclude` prop on `<Prefetch>` lets you prefetch a page's shell and cached content while skipping expensive dynamic sections, reducing bandwidth and latency on hover.
Drizzle ORM
1 update
8 Oct
Postgres-js driver now uses prepared queries with explicit .prepare()
The postgres-js driver was ignoring explicit `.prepare(name)` calls and not using prepared queries, which hurt performance and security. Now fixed.
A tool that ships more than twenty items in one edition has the rest follow in the next rather than being dropped.
This is what shipped each week, shaped by the same code that builds the email. A real edition also carries items an earlier one held back, skips anything you have already been sent, and never reaches back past the week you subscribed.