Vercel AI Updates: April 18, 2026

1. Zo Computer Cuts AI Retry Rates from 7.5% to 0.34% After Moving to AI SDK and Gateway

Vercel. Vercel shared a case study on Zo Computer, an 8-person personal-AI-cloud startup that migrated its multi-model orchestration to the AI SDK and AI Gateway. Retry rates dropped from 7.5% to 0.34%, chat success hit 99.93%, and P99 latency fell 38% from 131s to 81s across millions of daily model calls. The team also reports that new models now go live in 30 seconds instead of an hour because the Gateway handles provider-specific adapter code. Source

2. Deployment Retention Policies Now Preserve Active Branch Previews

Vercel. Retention policies will no longer delete the latest preview deployment for any branch that has an open pull request, giving teams a safer way to use short retention windows without breaking review flows. Vercel also preserves the 10 most recent production deployments alongside the policy so rollbacks stay available. Source

3. AI SDK Workflow v1.0.0-beta.26 Promotes experimental_telemetry to Stable

Vercel. The AI SDK Workflow package’s beta.26 release graduates experimental_telemetry to a stable API and refactors signal merging for cleaner cancellation semantics. The change lands alongside renamed tool-execution events shipped in beta.24 earlier in the day, pointing toward a nearer-term 1.0 cut for the workflow primitive. Source