Vercel AI Updates: May 13, 2026
1. AI Gateway exposes fast mode for Claude Opus 4.7
Vercel. The AI Gateway now surfaces Anthropic’s research-preview “fast mode” for Claude Opus 4.7, which Anthropic clocks at roughly 2.5x faster output token generation with no loss in model intelligence. Callers opt in by passing speed: 'fast' in the anthropic provider options, with pricing set at 6x standard Opus rates ($30/$150 per million input/output tokens). It is aimed at latency-sensitive agent loops where the standard Opus throughput becomes the bottleneck. Source
2. AI SDK gateway provider gains retry support for gateway errors
Vercel. @ai-sdk/gateway@4.0.0-canary.78 ships a fix that enables retry support for errors originating in the gateway itself, alongside updates to the gateway model settings surface. Until now, failures attributable to the gateway layer would bubble up without participating in the SDK’s standard retry policy, forcing callers to wrap calls in their own backoff. The canary brings gateway errors into the same retry budget as provider errors. Source
3. Mistral provider forwards stopSequences as native stop parameter
Vercel. @ai-sdk/mistral@4.0.0-canary.44 changes how the provider handles stopSequences, forwarding them through Mistral’s native stop parameter instead of relying on client-side post-processing. The practical effect is that stop strings actually halt generation at the model rather than truncating after the fact, which matters for both token billing and streaming UIs that no longer leak the stop string before cutting it. Source
4. AI SDK workflow canary 50 lands with dependency refresh
Vercel. @ai-sdk/workflow@1.0.0-canary.50 is a dependency-only bump that pulls in the latest ai and @ai-sdk/provider-utils canaries, with parallel releases across the Vue, Svelte, RSC, React, OTel, LlamaIndex, and LangChain packages on the same commit. No new workflow API surface today, but the coordinated cut keeps the canary track aligned after this week’s tool-description and step-numbering changes. Source