System OneSystem One
All articles

September 18, 2026 · 2 min read

Jev, explained — the first System One model

Jev is the model that named the category. What it takes in, what it returns, what it costs, and what its makers have not said.

Biplov Gautam

TypeSafe AI shipped Jev on 15 September 2026 and, with it, the phrase System One model: a model that makes fast, typed decisions instead of writing. The name comes from Kahneman's fast, intuitive System 1. Jev is closed, served only through TypeSafe's API, and the reference point every open model since has measured itself against.

What you send

One request carries a state — text, JSON or an array, up to 32k tokens — and a set of questions, each typed:

{
  "state": "Task: clean up inactive accounts before the quarterly report ...",
  "model": "jev-latest",
  "questions": {
    "safe": {
      "type": "noul",
      "instructions": "Is this action safe to run without a human approving it first?"
    },
    "queue": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": { "billing": null, "security": null, "support": null }
    },
    "urgency": {
      "type": "score",
      "instructions": "How urgent is this?",
      "criteria": ["Can wait a week", "This week", "Today", "Now"]
    }
  }
}

Choice takes up to 255 options; Score takes 2–10 levels; Noul takes a condition. The questions are answered in parallel in one pass.

What comes back

One answer per question, each carrying a probability distribution rather than a sentence:

{
  "model": "jev-1.13.0",
  "answers": {
    "safe": { "type": "noul", "noul": 0.07 },
    "queue": { "type": "choice", "choice": "security",
               "probabilities": { "billing": 0.02, "security": 0.91, "support": 0.07 },
               "confidence": 0.91 },
    "urgency": { "type": "score", "score": 2.8, "confidence": 0.72,
                 "probabilities": { "0": 0.0, "1": 0.05, "2": 0.15, "3": 0.8 } }
  },
  "usage": { "input_tokens": 412, "output_tokens": 0 }
}

Nothing is generated, so nothing has to be parsed and no field can be invented. TypeSafe describes the probabilities as calibrated by RLCD training on synthetic data; a 0.9 should be right about nine times in ten across similar cases.

Cost and speed

Input is $0.042 per million tokens and output is free — there is none. TypeSafe quotes 70–500 ms per request; independent measurements report a median of 236–276 ms. Rate limits at launch were 250k tokens per second and 1,200 requests per minute. The waitlist came down on 20 September.

What is not known

Jev has no published weights, parameter count or paper. Its accuracy is known only through other people's benchmarks: 76.0% macro on Bespoke Labs' thirteen public subsets, 93.2% on their held-out set, and the top score on JevBench. Those are the numbers the open families quote when they claim parity.

Where the open models stand

Because Jev fixed the request and response shape, almost every open System One model serves the same /v1/systemone contract. If you want the shape without the hosted dependency, Kev, Laya, Nimble and Decider all run on your own hardware under Apache-2.0. Jev's page on the registry records its pricing, status and API documentation alongside them, so the choice is one comparison away.