What Is Jev?
Jev is the first public model from TypeSafe AI, released on 15 September 2026. It is not a chatbot, and calling it an “LLM” will give you the wrong mental model. Jev does not write text. You send it a state — a message, a record, the current state of your application — together with one or more questions about that state, and it returns typed answers with probabilities attached.
The Short Answer
TypeSafe describes Jev as “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
That one line is the whole idea. An LLM asked to triage a support ticket will write you a paragraph that you then have to parse. Jev asked the same question returns:
{ "is_urgent": { "type": "noul", "noul": 0.999 } }That is a 99.9% probability the message is urgent. Your code branches on the number directly — no parsing, no format drift, no refusals.
What “System One” Means
TypeSafe positions System One as a new class of frontier model: built to make fast, structured decisions that software can use directly, rather than to converse with people. The name comes from the System 1 / System 2 idea in psychology — fast intuition versus deliberate reasoning; the company’s announcement links to Kahneman’s Thinking, Fast and Slow.
Two pieces of engineering make it possible, according to TypeSafe: a parallel sampler that produces every output in a single pass instead of one token at a time, and a training method the company calls Reinforcement Learning for Calibrated Decisions (RLCD). Where RLHF optimises for responses human raters prefer, and RLVR optimises for outputs that can be programmatically verified, RLCD optimises for honest probabilities on decision tasks.
TypeSafe was founded by Diogo Almeida, who worked at OpenAI on the methods behind ChatGPT. Jev came out of two years in stealth.
How Jev Differs From an LLM
| Existing LLMs | Jev (System One) | |
|---|---|---|
| Optimised with | RLHF / RLVR | RLCD |
| Output | Generated text — needs parsing and validation | Type-safe structured values, defined up front |
| Sampling | Sequential, one token at a time | Parallel, single pass |
| Input cost | $0.20–$10 per million tokens | $0.042 per million tokens |
| Output cost | ~5x the input cost | Free |
| End-to-end latency | 3–329 seconds | 70–500 ms |
| Confidence | Often overconfident and inconsistent | Calibrated on every output |
Figured from TypeSafe’s launch post. Two caveats the company states itself: the published speed evals were run from laptops on the US West Coast, where the service is currently hosted; and the headline numbers on their homepage (193.6x faster, 444.6x cheaper) come from their own workflow evals and are, in their words, likely “on the higher end of real world gains.”
The Three Question Types
Every question you can ask Jev is one of three primitives:
- Choice — pick from a set of options. Returns a probability per option plus an overall confidence score.
- Score — rate an input against ordered levels (calm / frustrated / very angry). Returns a continuous score, the underlying distribution, and a confidence value.
- Noul — a yes/no question. Returns the probability that the statement is true.
You can mix all three in a single request, and every question is evaluated in parallel — adding more questions barely changes the response time.
What Jev Can’t Do
Being precise here matters more than the launch hype:
- No text generation. Jev cannot write your reply, summarise a document, or answer in prose. It decides; something else writes.
- Text only. State must be a string, a JSON object, or an array of text values. Images, audio and video are not supported yet.
- English first. English is its primary training language. Other languages, including Chinese and Japanese, are accepted but currently score lower accuracy.
- Early access. You need to be granted access before you can call it.
- “No type errors” is a design guarantee, not a benchmark. Because the output schema is fixed in advance, an out-of-schema answer is mathematically impossible. A confident wrong label is still possible — which is exactly why every answer ships with a confidence score.
FAQ
Is Jev an LLM? No. It is a System One model: it returns typed decisions and probabilities rather than generated text. TypeSafe positions it as a complement to an LLM, not a replacement.
How much does Jev cost? $0.042 per million input tokens, and output tokens are free. TypeSafe notes it cannot prove the pricing isn’t subsidised, and says it expects pricing to go down over time, not up.
Can Jev hallucinate? It cannot return an answer outside the schema you defined. It can still return a label that is wrong but valid, so read the confidence value rather than trusting the top choice blindly.
Does Jev replace GPT or Claude? No. LangChain’s framing is the useful one: use an LLM for open-ended reasoning and generation, and Jev for the fast, structured decisions in between.
How do I get access? Request early access from TypeSafe, then start in the playground. See How to Use Jev for the step-by-step, or the Jev API guide for request and response details.