Jev decision-only model

Paid preview — This review is free to read. Future deep-dives like this one are for members only.

A model came out last month that cannot write a sentence. That’s not an insult — it’s the entire product thesis.

Jev, from the stealth lab TypeSafe AI (founded by former OpenAI researcher Diogo Almeida, $40M raised, two years in stealth), was released September 15 and fully opened September 21, 2026. The company calls it the first of a new category: “System One Models” — systems built solely for fast, low-cost judgments.

What it actually does

You hand Jev a blob of unstructured state plus a set of typed questions. It hands back typed answers. That’s the whole surface — three question types:

  • choice — pick one of up to 255 named options
  • score — land on an ordinal scale you define (e.g. calm → annoyed → angry → furious)
  • noul — a yes/no question returned as a probability

A real call looks like this: feed it a support ticket (“my card was charged twice…”), ask for department (choice), urgency (noul), and anger (score) — and get back {team: billing, urgent: 0.85, anger: 2.07} in about 400ms for a fraction of a cent. Note the anger score: 2.07, not 2. The model tells you the customer is a hair past angry and drifting toward furious. You can sort a queue by that. You can’t sort a queue by the word “angry”.

The benchmark data

Independent testers have been putting Jev through its paces, and the numbers are striking:

Accuracy. In a 19-question paired benchmark against Laya (421M), Jev got 92.9% of judgments right vs 65.3% for Laya, winning every question by 4.7–55.6 points. Its probabilities are usable as-is (Brier scores 0.002–0.024); Laya’s needed per-question tuning. On 128 non-English tickets across seven languages: Jev 100%.

Speed & cost. This is the headline. Official benchmarks claim Jev runs roughly 194× faster and costs about 1/445th as much as frontier generative models on suitable tasks, with output tokens free. One independent run of 40 support tickets found it 10× faster than Claude Fable 5.1 and 329× cheaper — while beating Fable at judging customer sentiment.

Adoption. Vercel reported nearly 13% of its paid AI Gateway teams tried Jev within the first day — higher day-one uptake than several recent large-model launches.

The catch

Jev is not a general model. It can’t write, can’t code, can’t chat — its 512-token-class constrained surface is the point, but also the boundary. The honest framing: it’s a specialist instrument, not a frontier competitor. For classification, routing, scoring, and triage at scale, it’s arguably the best price/performance in the industry right now. For everything else, you still need the big models.

Verdict

Jev is the clearest signal yet of a market split: generative models for open-ended work, decision models for judgments. If your pipeline parses LLM prose into structured fields today — and pays frontier prices for the privilege — Jev deserves a benchmark run this week.

Score: 9/10 for what it claims to be. The most interesting model release of the fall that isn’t trying to be AGI.

Sources: independent benchmark by Harry Munro; hands-on review at thoughts.jock.pl; TMTPost. Figures are as reported by these sources; your mileage may vary by workload.