Last updated: September 2026
Is Jev an LLM?
No. Jev is a System One model: it takes unstructured state plus typed questions and returns typed decisions with probabilities. It generates no text, so it cannot write a sentence, a summary, a JSON document or a line of code.
TypeSafe's own one-line description is "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Below: the technical difference, a side-by-side table, and why it changes what you build.
The technical difference is the decoding step
A language model produces its answer one token at a time, and each token conditions the next. That loop is where the latency comes from, where the cost comes from, and where a structured output can go off the rails. Constrained decoding does not remove the loop, it only restricts what the loop is allowed to emit.
Jev has no loop. All declared questions are evaluated in a single parallel pass, and each answer comes back as a probability distribution over the answer space you declared. Nothing is generated, so there is nothing to parse and nothing to retry. TypeSafe puts end-to-end response time at 70ms to 500ms.
The training objective differs too. TypeSafe calls its method RLCD, Reinforcement Learning for Calibrated Decisions, and says it optimises for answers with honest probabilities rather than for human preference. There is no published paper, and the CEO said on Hacker News that the architecture is close to the chest for now. Treat RLCD as a vendor description, not as a documented method.
// in
{
"state": "Shipped the rewrite. 40% fewer lines, same features.",
"model": "jev-1.13.0",
"questions": {
"hook": { "type": "score", "instructions": "How well does the opening line stop a scroll",
"criteria": ["No hook", "Functional", "Strong"] },
"is_bait": { "type": "noul", "instructions": "The post asks for likes, replies or follows" }
}
}
// out
{
"model": "jev-1.13.0",
"answers": {
"hook": { "type": "score", "score": 1.73, "legend": { "0": "No hook", "1": "Functional", "2": "Strong" }, "confidence": 0.78 },
"is_bait": { "type": "noul", "noul": 0.02 }
},
"usage": { "input_tokens": 58, "output_tokens": 4 }
}There is no place in that response for a sentence, because there is no generation step that could produce one.
Jev and a language model, side by side
Latency and price figures are TypeSafe's own, from the launch post and the models page, September 2026. The GPT-5.6 Terra list price is the comparison The Register used.
Not more accurate. Cheaper and faster at similar accuracy
On TypeSafe's own four-workflow eval, Jev scores 67.8% against GPT-5.6 Terra at 67.9%, with Sol at 74.1% and Opus 5 at 73.1% both ahead of it. The cost per case is $0.0004 against Terra's $0.0304, and the latency 0.4s against 10.1s. The eval has no human ground truth: reference answers are the average of two other frontier models. Full table in Jev vs GPT.Why the difference matters in practice
If Jev were an LLM you would choose between them. Because it is not, you use both, and the question becomes which part of your pipeline each one owns.
- Jev decides. Which bucket, which level, is this true. One request can carry every question you have about a piece of content.
- A language model writes. The reply, the draft, the summary. Anything whose output is words.
- Something publishes. The decision and the draft are worthless until the post exists. That is the part OpenTweet does, across X, Bluesky and LinkedIn.
The reason this ordering is worth caring about is cost per decision. A judgment that costs a fortieth of a cent can run on everything: every reply in a feed, every draft variant, every incoming DM. A judgment that costs three cents and takes ten seconds runs only on the items you already selected, which means something cruder did the selecting.
The split runs in production behind the free post scorer, which sends one Score for reach strength, three Scores for hook, clarity and specificity, and three Nouls for AI-slop, engagement bait and toxicity. Seven questions in one request, and no language model is involved, because nothing on that page has to be written.
Frequently asked questions
Is Jev an LLM?
No. Jev is a System One model. You send it state plus typed questions, and it returns typed decisions with probabilities TypeSafe describes as calibrated. It generates no text at all, so it cannot write a sentence, a summary, a JSON document or a line of code.
What is a System One model?
TypeSafe defines it as "a new class of frontier models built to make fast, structured decisions that software can use directly." The name is a nod to Kahneman's System 1, the fast intuitive mode of thinking, as opposed to System 2 deliberation. TypeSafe describes Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."
Can Jev replace my LLM?
Only for the decisions. Anything that has to produce words, code or structured documents still needs a language model. Jev covers the judgment calls a pipeline makes on the way: which bucket, which level, is this true. Most production setups use both.
How is Jev different from OpenAI structured outputs?
Structured outputs constrain a language model's token-by-token decoding so the text it generates parses as your schema. Jev never generates tokens: all declared questions are evaluated in one parallel pass and the answer comes back as a probability distribution over your declared answer space. That is why TypeSafe puts its end-to-end response time at 70ms to 500ms rather than seconds.
What was Jev trained with, if not RLHF?
TypeSafe calls the method RLCD, Reinforcement Learning for Calibrated Decisions. It says RLCD optimises for "answers with epistemically honest probabilities on System One tasks" rather than for human preference. No paper has been published and TypeSafe's CEO said the architecture is close to the chest for now.
Keep exploring
The primitives, the limits, and the tool Jev runs behind.
Can Jev hallucinate?
Schema conformance is guaranteed. Factual accuracy is not, and TypeSafe conceded it.
Choice vs Score vs Noul
The three question types, with the exact request and response JSON for each.
Jev vs GPT
The four-workflow benchmark as a table, and what each model cannot do.
How to get Jev API access
The waitlist, and the three gateways that skip it.
Will it go viral?
A free X post scorer running on Jev in production. No signup.
Social media API for AI agents
The publishing layer an agent calls once the decision is made.
Watch a System One model make a call
Will it go viral? scores any X draft with seven Jev questions and shows the confidence band, free and with no signup. When the decision is made and the draft is written, the OpenTweet API is the one call that publishes it.