npm / @openrouter/ai-sdk-provider
@openrouter/ai-sdk-provider
What changed in @openrouter/ai-sdk-provider 2 to 3, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-19.
2 breaking2 security10 new APIs and options1 behavior and defaults16 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli init@openrouter/ai-sdk-provider 34 changes
Breaking changes 1
Image generation switches to a dedicated images API.
New APIs and options 2
Adds evaluationModel backed by the Decisions API.
The provider adds support for Vercel AI SDK v7.
Fixes and other changes 1
Consecutive streaming reasoning.summary deltas are now merged.
@openrouter/ai-sdk-provider 227 changes
Breaking changes 1
Accumulated reasoning_details are no longer attached to reasoning-delta events.
Security 2
Signature validation now also covers the Gemini format.
Entries without valid signatures are stripped from reasoning.text.
New APIs and options 8
You can now opt out of response_format strict mode.
Adds video generation support through VideoModelV3.
Anthropic eager_input_streaming is now supported via tool providerOptions.
Temperature and other common settings can now be set in model-level configuration.
Adds a provider-defined web search tool for use with OpenRouter.
Adds support for video URLs.
Adds the appName option for app identification headers.
Adds the appUrl option for app identification headers.
Behavior and default changes 1
doGenerate now includes the raw response body.
Fixes and other changes 15
Sends null content for tool-only assistant messages.
Trailing-whitespace deltas no longer emit duplicate tool-call events.
The image URL regex now allows query strings and fragments.
Preserves empty reasoning_details arrays in multi-turn conversations.
A defensive usage fallback was added in the streaming flush handler.
Serializes tool call arguments deterministically for cache consistency.
Enforces unique tool call IDs for parallel tool calls.
Streaming tool call lifecycles now emit tool-input-end events.
The provider now respects AI_SDK_LOG_WARNINGS when controlling warning logs.
Prevents duplicate reasoning blocks when a signature arrives after text.
Reasoning and text streams now use unique IDs.
Stops emitting [REDACTED] for encrypted reasoning details.
Infers the tool-calls finish reason when finish_reason is unknown.
Handles socket termination during a stream gracefully.
Tool messages now include the tool name to match the OpenAI API specification.