Vercel AI Updates: May 15, 2026

1. AI SDK xAI provider adds ‘none’ and ‘medium’ reasoning effort for grok-4.3

Vercel. @ai-sdk/xai@4.0.0-canary.61 extends the provider’s reasoning-effort surface so callers can set 'none' (disabling thinking tokens entirely) or 'medium' (extra thinking budget for less time-sensitive paths) on grok-4.3. The top-level reasoning: 'medium' parameter now maps correctly to reasoning_effort: 'medium' on the chat model instead of being coerced to 'low', and the model-ID autocomplete list is curated to grok-4.20-non-reasoning, grok-4.20-reasoning, grok-4.3, and grok-latest while the type stays open for any xAI-supported ID. Source

2. xAI provider stable line backports the additionalProperties flag fix

Vercel. @ai-sdk/xai@3.0.90 ships the stable-channel counterpart of the canary patch that stops emitting an extraneous additionalProperties flag in the xAI output schema. If you saw schema-validation errors on grok tool calls from the stable 3.x line this week, this is the patch that lands the fix without forcing a jump to the 4.0 canary track. Source

3. Mistral provider stable 3.0.37 forwards stopSequences as the native stop parameter

Vercel. @ai-sdk/mistral@3.0.37 backports the canary-track change so the stable 3.x line forwards stopSequences through Mistral’s native stop parameter rather than relying on client-side post-processing. The practical impact is that stop strings now halt generation at the model, saving billed output tokens and cleaning up streaming UIs that previously leaked the stop string before truncating it. Source