TypeSafe AI's Jev brings typed probabilistic decisions to Vercel's AI Gateway

Jev, a probabilistic decision model from TypeSafe AI, is now available through Vercel's AI Gateway, according to the Vercel blog. Instead of generating text that an application must parse and validate, Jev takes a state as input and returns typed Choice, Score and Boolean answers with probabilities attached.

A decision model instead of a text generator

The Vercel blog frames the difference this way: ordinary language models emit text one token at a time, and it is then up to the calling application to parse, validate and act on that output. Jev evaluates all declared questions in parallel and returns typed answers directly, which the blog says eliminates unnecessary text generation, lets clear cases be automated, and makes it straightforward to route uncertain ones to review.

Suggested use cases include picking the next tool or subagent in an agent loop, deciding whether to continue, retry, ask the user or stop, scoring urgency or risk before an action, and verifying model outputs to enforce guardrails.

Vendor-reported performance numbers

TypeSafe reports that Jev was up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations. These are the vendor's own benchmarks on its own workloads, and the "up to" framing means typical results may be lower; the blog offers no independent comparison.

Access via the experimental evaluate API in AI SDK 7

Jev is exposed through the experimental evaluate API in AI SDK 7, with support arriving in version 7.0.105. Three question types are available: Choice selects among options, Score grades against an ordered rubric, and Boolean estimates the probability that a statement is true.

Each evaluation defines the model to call (typesafe-ai/jev on the Gateway), the state to evaluate — a shared string, object or array — and a map of named questions about that state. Vercel's worked example feeds a one-line summary of a support interaction into a single Boolean question asking whether a refund was issued, then reads the answer from result.answers.refunded.

Answers preserve question IDs and Choice keys, and TypeSafe provides separate confidence values for Choice and Score answers through result.providerMetadata.typesafe.confidence. The blog recommends calibrating both probabilities and confidence against labeled examples from your own workflow before relying on them for production decisions.

Data handling and Gateway integration

Jev supports Zero Data Retention and No Training modes, enabled per request in the provider options. Evaluation calls appear in Gateway logs and custom reporting, count toward budgets, and accept the same providerOptions.gateway settings used with other Gateway providers, so teams can slot Jev into existing spend and audit setups without new tooling.

Why it matters

Much of the plumbing in modern agent applications exists to bridge the gap between a model's text output and the decisions a program actually needs: structured-output prompts, parsers, validators and retry logic. A model class designed to return typed decisions with probabilities collapses that bridge, at least for the routing and gating steps that would otherwise burn full LLM calls.

The economics matter as well. Agent loops make many small, frequent decisions, and paying standard LLM prices for each one adds up quickly; a purpose-built decision model with the reported cost profile could make longer, more autonomous loops viable. The caveats are real — the API is explicitly experimental, the performance figures come from the vendor, and probabilities still require calibration against real labeled data — but the arrival of decision models as a distinct category on a major gateway points toward agent architectures that match model types to tasks rather than pushing everything through a single general-purpose model.