Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 293 Agent status: 2 with issues Updated Oct 7, 2026
AgentGid

AI SDK changelog: what's new each month

Every stable AI SDK release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 2 months covered; the current month updates daily.

October 2026 · so far

7 releases: 7.0.127 → 7.0.130

October brought refinements to tool handling, reasoning capabilities, and stream management across the AI SDK. New features focused on tool search flexibility, portable reasoning levels, and improved error handling.

Action needed
  • Evaluate renamed to decide.
  • Amazon Bedrock Nova 2 with portable reasoning xhigh now sends maxReasoningEffort: 'high' instead of unsupported 'max'.

Highlights

  • Tool search now supports custom selection and ranking via a search() callback with configurable result limits.
  • Added portable max reasoning level with native provider mappings and compatibility fallbacks across Anthropic, Amazon Bedrock, Fireworks, and other providers.
  • Introduced durable non-streaming WorkflowAgent.generate() for workflows requiring typed output and core generation result semantics.
  • Renamed evaluate to decide for consistency across the SDK.

New

  • Tool search callback for custom selection and ranking of eligible deferred tools.
  • Configurable maxResults option for limiting tools returned in tool search.
  • Custom reasoning delimiters support in extractReasoningMiddleware.
  • Portable max reasoning level with provider-specific mappings and budget-based fallback.
  • Durable non-streaming WorkflowAgent.generate() with 20-step default and shared tool loop.
  • Deprecation warnings with stable codes and deduplication.
  • Decision refusals reported as refusal answers from providers.

Improved

  • Tool approval state now preserved when resuming streams.
  • Provider options preserved from approval-only tool messages.
  • Stream resumption now publishes the latest error after repeated failed attempts.
  • Silent audio now returns successful empty transcripts instead of errors.
  • Tool approval inputs from different JavaScript realms are now accepted.
  • V2 provider-reported total tokens preserved in raw usage.

Fixed

  • Merged UI message streams now cancel when the consumer disconnects.
  • Pending result promises reject and pending streams cancel when abort signal fires.
  • Invalid audio file format errors now throw InvalidResponseDataError for better error identification.
  • Exported isDynamicToolUIPart from ai package.

September 2026

8 releases: 6.0.297 → 7.0.126

September brought stability improvements and security enhancements across the SDK, with a focus on fixing stream handling, tool execution, and connection security. The month also introduced telemetry for speech operations.

Highlights

  • Stream resumption now correctly preserves partial tool input and keeps idle streams alive with optional heartbeats.
  • Chat IDs in reconnection URLs are properly encoded to prevent path confusion, with security checks for malformed IDs.
  • Tool execution errors now handled more safely, preserving provider error data while maintaining stream integrity.
  • Speech generation and transcription now include telemetry tracking.
  • Enhanced Node.js download security with DNS validation and connection pinning against framework wrapping.

New

  • Telemetry added to speech generation and transcription operations.

Improved

  • Stream resumption preserves hydrated partial static tool input across reconnections.
  • Idle UI message streams stay open longer with optional SSE heartbeats.
  • Chat IDs in default reconnection URLs are encoded to preserve special characters.
  • Tool, provider, and provider metadata names are preserved as own properties to prevent prototype pollution.
  • Automatic chat resumption logic refined to avoid resuming after terminal text without completed stream state.
  • Tool content pruning improved when retaining zero trailing messages.
  • Partial reasoning tags now preserved when streamed text parts end.
  • Pending tool approvals superseded by user messages are now properly ignored.

Fixed

  • Stream resumption now preserves hydrated partial static tool input correctly.
  • Idle UI message streams no longer close unexpectedly.
  • Chat ID encoding prevents slashes and query delimiters from breaking reconnection URLs.
  • Prototype-named tools and providers no longer cause inheritance resolution issues.
  • Automatic chat resumption no longer triggers incorrectly after terminal text.
  • Tool approvals are now properly cleared when addToolOutput runs.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. AI SDK: pricing, features and alternatives · All changelogs