TypeSafe Jev AI

TypeSafe Jev, in one paragraph: TypeSafe AI launched Jev on 15 September 2026 after two years in stealth. The founder, Diogo Almeida, previously worked on RLHF methods behind InstructGPT and ChatGPT. Jev is the company’s flagship model and the first public System One model — a class of models trained to make fast, structured decisions that software can use directly. You send state (a string, object, or array) and a map of typed questions. Jev returns answers under the same keys — a Choice, a Score, or a Noul — with probabilities. It does not generate text. There is nothing to parse.

If you asked “what is Jev” because a blog said it is 200× faster than a frontier LLM, the vendor’s own comparison is on System One tasks, not on writing a novel. What TypeSafe Jev is good at is the decision inside the request path: route, score, gate. What it is bad at is language. Jev Patterns only publishes jobs in the first bucket.

Why “System One”

Kahneman’s System One is fast, associative judgment. TypeSafe uses the name for models that skip chain-of-thought prose and emit a decision your code can branch on. They also describe a training method, Reinforcement Learning for Calibrated Decisions (RLCD), as the counterpart to RLHF’s preference-for-prose. You do not need to believe the branding to use the API. You do need to believe the contract: typed answers, closed criteria, probabilities you can threshold.

Default input is text, JSON objects, and arrays of text. Audio and video are not supported. Experimental image input is available on both Jev 1.13 and 1.14 behind the TypeSafe-Danger-Zone-Images: true header. The Box primitive (bounding box top/bottom/left/right as fractions 0–1 plus confidence; always returns a box — pair with a Noul for presence) is 1.14-only (jev-preview / jev-1.14.0). Longer context (~262k) is behind TypeSafe-Danger-Zone-Long-Context: true. TypeSafe marks these danger-zone features as dreamer-ready / not yet production-grade; format and limits can change. See docs.typesafe.ai/images and docs.typesafe.ai/primitives/box. Default context remains 64k per-request / 32k per-question. Re-read docs.typesafe.ai/models if you are near the limit or pinning versions.

Latency and price (vendor figures)

TypeSafe’s public figures at launch: end-to-end response time about 70–500ms. Input about $0.042 per million tokens ($42 per billion). Output unmetered, because Jev is not sampling a reply. Questions in one request are evaluated in parallel against the same state, so adding a Noul barely moves the clock. Rate limits documented for jev-1.13.0 include 1,200 requests per minute and a very large token-per-second cap — confirm on the models page before you plan capacity.

The public alias is jev-latest (currently pointing at jev-1.13.0; scheduled to move to jev-1.14.0 about a week after the October 14 release). jev-preview points at jev-1.14.0. The HTTP route is POST https://api.typesafe.ai/v1/systemone. Official product: typesafe.ai. Aliases move when a release ships; pin a versioned id if a threshold is a deploy gate.

Named after Jevons

Jevons’ paradox: as a resource gets cheaper to use, total use often rises. Cheap, calibrated decisions will be called in places nobody would have spent a frontier LLM turn. That is the point of a library of schemas — the decision is cheap enough to put on every ticket, every chunk, every agent step. What TypeSafe Jev is for, operationally: high-frequency judgment with a closed answer space.

What TypeSafe Jev is not

Not a chatbot. Not JSON mode on a writer. Not function calling. Those still sample tokens, even when a grammar keeps the braces valid. See Jev vs ChatGPT, Jev vs JSON, and Jev vs function calling. If the user needs a paragraph, you are in the wrong building — when not to use Jev.

A minimal TypeSafe Jev request

State can be as small as a ticket string. Questions can be a single Choice. The response keys match the question keys. That is the whole TypeSafe Jev contract at the wire:

{
  "model": "jev-latest",
  "state": { "ticket": "Password reset email never arrives" },
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which queue owns this?",
      "criteria": {
        "account": "Login, identity, reset mail",
        "billing": "Charges and invoices",
        "technical": "Product bugs and outages",
        "other": "None of the above"
      }
    }
  }
}

There is no messages array. There is no system prompt asking Jev to be a helpful assistant. TypeSafe Jev, mechanically: an evaluator of state against typed questions. If you find yourself adding “explain your reasoning,” you have left the product. Put the explanation in your UI after you branch on department.choice and department.confidence.

TypeSafe Jev in an existing stack

Beside a helpdesk: classify and score, then let macros or humans talk. Beside RAG: score chunks and gate drafts. Beside an agent: pick the next tool and whether to stop. Beside a form: intent Choice, then CRM. What Jev is not in those stacks: the thing the user reads. Official SDK notes (Netlify AI Gateway and TypeSafe’s own clients) default to jev-latest. Jev Patterns schemas are the questions you pass into those clients. Still independent. Still not affiliated with TypeSafe AI.

Jev Patterns is independent and not affiliated with TypeSafe AI. We collect reusable request shapes. We do not host the model. Read what is Noul in Jev next, or skip to the catalog, or try the Jev playground. Need a key? Where to call Jev.