What Is Jev? TypeSafe's System One Model Explained

Jev is the first model in what TypeSafe AI calls the System One family. You give it some content (the state) and a set of questions whose possible answers you define in advance. Jev returns an answer to every question, each with probabilities attached. It never writes a sentence, a reply or a piece of code.

That makes it a different kind of tool from ChatGPT, Claude or Gemini. An LLM is built to produce text for people to read. Jev is built to produce values your program can use directly in an if statement, a routing table or a ranking.

What a Jev call looks like

Here is a real request we sent on September 19, 2026, pinned to the current model version jev-1.13.0. The state is one customer message; the three questions use Jev’s three question types.

Request (abridged)

{
  "model": "jev-1.13.0",
  "state": "My invoice lists two seats, but only one of us can sign in, and the login page keeps timing out.",
  "questions": {
    "queue": {
      "type": "choice",
      "instructions": "Which team should handle this message?",
      "criteria": {
        "billing": "Charges, invoices, refunds, seats on the bill",
        "technical": "Bugs, outages, login problems, integrations",
        "sales": "New plans, upgrades, quotes",
        "other": "Anything else"
      }
    }
  }
}

Jev answered in 586 ms with "choice": "technical", a confidence of 0.94 and this probability for each option:

technical0.96

billing0.04

sales0.00

other0.00

Probabilities returned by jev-1.13.0 for the message above (our test, Sep 19, 2026).

The answer can only ever be one of the four keys you supplied. There is no free text to parse, and the probabilities tell your code how sure the model is.

The three question types

Type	You ask	Jev returns	Typical use
Noul	A yes/no question	noul: probability of “yes”, 0 to 1	Filters, guardrails, “is this spam?”
Choice	Pick one option from a list (up to 255)	choice, a probability per option, confidence	Routing, classification, picking the next action
Score	Place the state on levels you describe	score (can fall between levels), probabilities, confidence	Severity, urgency, quality ratings

You can mix all three in one request and ask many questions at once. Every question sees the same state and is answered independently.

How Jev differs from an LLM

Aspect	Jev (System One)	Chat LLMs
Output	Typed values you define in advance, with probabilities	Free text
Training target	Calibrated decisions (TypeSafe calls this RLCD)	Human preference (RLHF) or verifiable rewards (RLVR)
Output tokens	Free	Usually several times the input price
Typical latency	70 to 500 ms according to TypeSafe	Seconds to minutes
Can it write an email or code?	No	Yes

TypeSafe’s founder describes the idea as a “frontier-intelligence function call”. The blog post that launched Jev claims it matches existing LLMs on these narrow decision tasks while being about two orders of magnitude faster and cheaper. Those comparisons come from TypeSafe’s own workflow evaluations; see Jev benchmarks explained for what they measure.

Key numbers (as of September 19, 2026)

Item	Value
Current model	Jev 1.13 (jev-1.13.0); the aliases jev-latest and jev-preview both point to it
Price	$0.042 per million input tokens ($42 per billion); output tokens are free
Context	64k tokens per request; 32k for the state plus the longest single question
Rate limits	1,200 requests per minute and 250,000 tokens per second, which TypeSafe says are adjusted dynamically
Input	Text only: a string, a JSON object or an array. No images, audio or video
Languages	English is most accurate; other languages work with lower accuracy

In our own tests from East Asia, ten small calls in a row took between 253 and 378 ms, with a median of 284 ms. Asking 20 questions in one request took no longer than asking one.

Who makes Jev

Jev is made by TypeSafe AI, a San Francisco startup led by Diogo Almeida, a former OpenAI researcher who co-invented RLHF and worked on InstructGPT. The model is named after the economist William Stanley Jevons, whose paradox says that when something gets cheaper people use far more of it. The “System One” label borrows Daniel Kahneman’s term for fast, intuitive thinking. More on the company: TypeSafe AI, the company behind Jev.

How to use it

You can call Jev on TypeSafe’s own API (early access), or through Vercel AI Gateway, OpenRouter and Cloudflare Workers AI, which each list it with slightly different model names and limits. How to get a Jev API key walks through each route, and Jev channels compared puts them side by side.

What it is good at, and what it is not

Jev fits decisions that a person could make at a glance: which team gets a ticket, whether a message is abusive, how urgent something is, or which of 182 skills an agent should load (one of TypeSafe’s own examples). Developers have used it to route prompts between models, drive browser agents one click at a time and check LLM output for jailbreaks.

It is not a replacement for an LLM when you need text. TypeSafe’s own list of known weak spots for Jev 1.13 includes arithmetic, counting, comparing dates, multi-step reasoning and very long states full of irrelevant detail. See Jev limitations before you rely on it.

What developers are saying

The launch post reached nearly 1,900 points and about 500 comments on Hacker News within days. The most common questions were whether Jev is an LLM at all, whether “it can’t hallucinate” is a fair claim, what the probabilities actually mean, and whether an open-weights version exists. We answer each of them in the FAQ: Is Jev an LLM?, Can Jev hallucinate? and Is Jev open source?.

Quick answers

Is Jev free? Not on TypeSafe’s own API, but Vercel AI Gateway was listing it as free under a promotion ending September 25, 2026. See Is Jev free?

Can Jev generate text? No. It only returns the typed answers you define. See Can Jev generate text?

What does it cost to run? At $0.042 per million input tokens, a typical request with a short message and a few questions (300 to 450 input tokens in our tests) costs about $0.00002. See Jev pricing.