What Is Jev AI? TypeSafe's System One Model Explained

Jev in one paragraph

Jev is the first model from TypeSafe AI, a San Francisco startup that launched on September 15, 2026. Most AI models are built to produce text for people to read. Jev is built to produce decisions for software to act on. You send it a state (a string or JSON describing the situation) and a set of typed questions. It returns one typed answer per question, with probabilities, in a single fast pass. There is no prose to parse and no chance of a malformed JSON response.

TypeSafe calls this a System One model, after Daniel Kahneman’s fast, intuitive “System 1” thinking. It is meant for the quick judgment calls that come up constantly inside software: which team should get this ticket, how urgent is it, is this message spam, which button should the agent click next.

How a Jev call works

Every call has the same shape: one state, many questions.

{
  "model": "jev-latest",
  "state": "Arrived two days late and the box was crushed, but the headphones sound great.",
  "questions": {
    "topic": {
      "type": "choice",
      "instructions": "What is the review mainly about?",
      "criteria": {
        "product": "The product itself",
        "delivery": "Shipping speed or packaging",
        "support": "Customer service"
      }
    },
    "needs_follow_up": {
      "type": "noul",
      "instructions": "The customer reports a problem the store should follow up on"
    }
  }
}

Each question is evaluated independently and in parallel against the same state, so adding a fifth or tenth question adds little latency and doesn’t dilute the others. The response contains one answer per question name, plus a probability distribution you can inspect.

The three question types

Type	Asks	Returns	Example
Choice	Which of these options fits?	choice, probabilities per option, confidence	Route a ticket to billing, technical, or sales
Score	Where does this sit on a scale you define?	score (can fall between levels), probabilities per level, confidence	Rate frustration from calm to very angry
Noul	Is this statement true?	noul, the probability of “yes” from 0 to 1	Does this message request a refund?

A Choice question accepts up to 255 options. A Score’s score is the probability-weighted average of your levels, so on a three-level scale it runs from 0 to 2 and can land at, say, 1.4.

Confidence is the point

Every Choice and Score answer carries a confidence value derived from how concentrated its probability distribution is. That lets your code act differently depending on how sure the model is:

• High confidence: act automatically.
• Medium confidence: ask for confirmation or flag for review.
• Low confidence: route to a person or a stronger model.

TypeSafe trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD) so that these probabilities reflect real uncertainty. Calibration is a property measured across many predictions. It doesn’t guarantee that any single answer is right, and the right thresholds depend on your data. Our getting started guide shows the basic pattern.

What Jev is good at, and what it isn’t

Jev does best on common-sense judgments a knowledgeable person could make in a few seconds. TypeSafe’s own Jev 1.13 jaggedness page is unusually candid about the rest. Jev 1.13 is weak at:

• Math, counting, and numeric precision. Do arithmetic in code.
• Comparing dates. Extract date parts with Jev if needed, then compare in code.
• Multi-hop reasoning and double negatives. Ask direct questions instead.
• Large states full of irrelevant detail. Filter first and send only what the question needs.
• Adversarial input. Text written to steer the model can move its answers.
• Generating text. Use a generative model.

It also reads instructions literally. If you find yourself explaining what a question “really meant,” that explanation belongs in the question.

The practical rule is to decompose. Instead of one broad question like “rate this startup pitch,” ask separate questions about market, feasibility, and differentiation, then combine the answers with your own formula in code.

“Can’t hallucinate”: what that claim does and doesn’t mean

TypeSafe’s launch post says Jev “can’t hallucinate.” That’s true in a narrow sense: Jev can only return values from the answer space you define, so it can’t invent an option, produce broken JSON, or make up a citation. It can still choose the wrong option, and it can be wrong with high confidence. Independent analysts, including Anthony Maio, have pointed out that constraining the shape of the output doesn’t constrain the judgment. Treat Jev like any other model: measure its accuracy on your task before automating on it.

Speed and price

Jev costs $0.042 per million input tokens, and output is free. TypeSafe reports responses in roughly 70 to 500 milliseconds. Its headline comparisons (“193.6× faster, 444.6× cheaper”) come from TypeSafe’s own published workflow evaluations, and TypeSafe itself says those numbers are likely at the high end of real-world gains. How much you save depends on your input size, how many output tokens your LLM alternative would generate, and which model you compare against. Try it with your own numbers in the cost calculator, or read Jev pricing explained.

Who makes Jev

TypeSafe AI was founded in 2024 by Diogo Almeida (previously a researcher at OpenAI), Erik Gafni, and Sasha Sheng. It came out of stealth on September 15, 2026 with a $40 million seed round led by DCVC. The current model is jev-1.13.0, reachable through the jev-latest alias.

How to try it

Access is through an early-access waitlist at typesafe.ai. Once you’re in, you can experiment in the browser playground, or call the API from Python, JavaScript, or plain HTTP. Our getting started guide walks through every step, and the project directory shows what others have already built.

Frequently asked questions

Is Jev an LLM?

Not in the usual sense. Like an LLM it understands natural-language input, but it never generates text. It returns typed answers (a choice from your options, a position on your scale, or a yes/no probability). TypeSafe calls this class of model a System One model.

Can Jev write text or code?

No. TypeSafe's docs say Jev is not trained to generate text and recommend using a generative model for that. If you need to extract a value, the recommended approach is to generate candidates another way and let Jev choose among them.

How much does Jev cost?

$0.042 per million input tokens ($42 per billion). Output tokens are not billed. See our pricing page and cost calculator for worked examples.

Does Jev hallucinate?

Jev can't produce malformed output or invent an option you didn't offer, because it only returns values from the answer space you define. It can still pick the wrong option or be confidently wrong, so validate it on your own data.

How fast is Jev?

TypeSafe reports end-to-end responses of roughly 70 to 500 milliseconds, and asking more questions in the same call barely changes response time because they are evaluated in parallel.

Is Jev open source?

No. The weights are not public. Several community projects reproduce the same interface on open models; see our Open Models & Reproductions category.

What languages does Jev support?

It accepts natural-language text. English is its primary training language and where accuracy is best. Other languages, including Chinese, Japanese, and Korean, work but less well, so test on your own content.

What is Jev's context limit?

64k tokens per request for the state plus all questions, and 32k tokens for the state plus the single longest question. Input must be text: a string, a JSON object, or an array of text values.