Sign in
← AI PulseSep 16, 2026

Wire · news · Multi-source brief

Vercel AI SDK Adds Experimental Evaluation Model Aliases and TypeSafe Provider

The Vercel AI SDK has introduced experimental evaluation model aliases and registry resolution, alongside a new TypeSafe provider for experimental_evaluate that supports native question types and structured JSON rubrics.

By Illumora Editorial

Source · Sep 16, 2026, 9:31 PM · On Illumora · Sep 16, 2026, 9:43 PM

Media from the primary source — shown here so you can stay on Illumora.

Synthesized from multiple allowlisted primaries on the same event. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →Vercel AI SDK Changelog — Release @ai-sdk/workflow@2.0.35 · vercel/ai
Save

The Vercel AI SDK has released updates including ai@7.0.104 and @ai-sdk/typesafe-ai@3.0.0. These releases introduce experimental evaluation model aliases and registry resolution, allowing customProvider to accept evaluationModels and registries to expose evaluationModel. The experimental_evaluate function now accepts string IDs when an evaluation-capable default provider is explicitly configured.

Additionally, a new TypeSafe provider has been added for experimental_evaluate, supporting native Choice, Score, and Boolean questions within a single request. This provider also enables structured JSON rubrics, probability distributions, confidence metadata, and workflow serialization, according to the Vercel AI SDK Changelog.

Key Points

  • The Vercel AI SDK now includes experimental evaluation model aliases and registry resolution in ai@7.0.104.
  • customProvider can accept evaluationModels, and registries can expose evaluationModel.
  • experimental_evaluate now accepts string IDs when an evaluation-capable default provider is explicitly configured.
  • Evaluation will not implicitly fall back to Gateway.
  • Model-resolution errors now identify evaluationModel.
  • The TypeSafe provider for experimental_evaluate supports native Choice, Score, and Boolean questions in one request.
  • This TypeSafe provider also supports structured JSON rubrics, probability distributions, confidence metadata, and workflow serialization.

Context

According to the Vercel AI SDK Changelog, these updates also include fixes for abnormal realtime WebSocket close diagnostics and support for tool search with direct tool calling. The ai@7.0.103 release, preceding ai@7.0.104, added mid-conversation tool discovery/updates for code-mode and protected default Node.js downloads with DNS validation and connection pinning.

Why It Matters

These updates provide developers with more granular control over model evaluation within the Vercel AI SDK, offering explicit configuration for evaluation models and a structured approach to assessing model outputs through the TypeSafe provider. This can lead to more robust and transparent evaluation workflows.

What To Do

  • Review the Vercel AI SDK documentation for details on experimental_evaluate and evaluationModels.
  • Explore the new TypeSafe provider to implement native Choice, Score, and Boolean questions in evaluation requests.
  • Investigate the use of structured JSON rubrics and confidence metadata for enhanced evaluation.
  • Note that evaluation will not implicitly fall back to Gateway, requiring explicit configuration.