Jev returns a decision, not a paragraph, and 13% of Vercel's paid teams grabbed it in a day

A new model showed up in September, and it does not talk. You hand it text and a fixed set of choices, and it hands back a decision your code can act on. No paragraphs, no chat window. Product teams are already wiring it into routing, triage, and interfaces. Let me catch you up on what changed and where this fits in your stack.

The deep cut

• A decision is a data type, not a paragraph. Jev returns a typed choice, score, or probability your code branches on, no text parsing.
• Confidence is the feature you route on. Bryo AI's CTO called Jev "the only one that hands back a real probability," useful because software can escalate coin-toss cases.
• Cheap intelligence gets used more, not less. Jev is named for Jevons, and at $0.042 per million input tokens teams are sprinkling judgment everywhere.

What "returns a decision" actually means

Jev comes from TypeSafe AI, founded by a ChatGPT co-inventor, Diogo Almeida, who helped build the RLHF technique behind modern chatbots. His complaint is plain: "We have lightning in a bottle, and yet it is not useful." Chat models speak human language. Software needs decisions.

So Jev does one thing. You give it a state and typed questions. It answers in parallel and returns values. Three kinds: Choice picks one option from a list, Score rates against a rubric, and Noul gives the probability a statement is true. Think of a smart if-statement. Your code can check order.total > 100 on its own. It cannot check "is this customer angry?" Jev handles that and hands back a number.

The confidence number does the real work

The output that matters is not the label, it is the certainty attached to it. In one hands-on test on support tickets, a clear refund demand came back 99% true and routed to billing at 100%. Strip out the anger and the demand, and refund dropped to 6%. On a genuinely fuzzy ticket, the model split 58% technical, 42% billing, with confidence at 0.44.

That split is the point. TypeSafe trains for calibration with a method it calls RLCD, so a 0.95 and a 0.50 mean different things. Your software auto-acts above a threshold and sends the coin-toss cases to a human. Almeida's own framing: at 50% "maybe this is a coin toss," at 95% "sure, then I can do something with it." That gate is a design decision, not a model setting.

Fast enough to change the interface

Speed opens uses a chatbot cannot touch. A Vercel Labs demo rendered the same travel ticket two ways: the LLM path took 3.68 seconds, the Jev path 0.88. Jev did not design a better ticket. It picked and arranged approved UI blocks faster. When a model chooses which components belong in a moment, your design system stops being a library and starts being the menu Jev picks from.

Adoption backs the interest. Jev hit nearly 13% of paid teams on Vercel AI Gateway in 24 hours, which Vercel called its fastest-adopted launch. One builder scored a newsletter draft for $0.000126 a request.

Where the hype outruns the facts

Before you rip out an LLM call, note the caveats. "Cannot hallucinate" means it cannot return a label outside your schema, not that it is right. Give it Billing, Technical, Sales, and it will never invent Legal. It can still pick Billing when the answer was Technical.

The accuracy numbers are thin. TypeSafe reports about 68% on its own evaluation, and that score measures agreement with two frontier models, not correctness against real labels. One ML practitioner who has built classifiers for years put it flatly: the problem is old, zero-shot classification has done this since 2019, but the architecture and calibration around it may be new. Jev is a good product bet, not a new form of AI. Treat the vendor benchmarks as promising, not proven.

Three questions for your team

• Where in our flows are we paying a frontier LLM to make a small yes/no or which-bucket call that Jev could do faster and cheaper?
• What confidence threshold do we set for auto-action versus human review, and who owns that number when it is wrong?
• If a model starts choosing which UI components render, is our design system defined tightly enough to be a safe menu, or too loose to trust?