Jev is a System One model, not an LLM
Jev is a System One model from TypeSafe AI, not an LLM. It returns a typed value and a calibrated probability from program state. It does not generate text.
Jev is not an LLM. It does not generate text. It takes unstructured program state and returns typed structured values, plus calibrated probabilities and confidence scores, in a single parallel pass. TypeSafe AI calls this class a System One model: fast focused judgements, named after the fast/slow split in Thinking, Fast and Slow.
TypeSafe's definition is in Introducing System One models and Jev. Limited early access is from 15 September 2026. This page is what a System One model is, and why Jev returns a typed value instead of a sentence. It is not a reading of the zero-hallucination claim, a content-policy routing design, or the cost arithmetic for a model router.
What a System One model is
Jev is the model. TypeSafe AI is the company: San Francisco, founded in 2024. Jev was announced alongside a $40M seed led by DCVC. Founder Diogo Almeida is ex-OpenAI. He helped build ChatGPT and worked on RLHF.
Alongside that post, the public pages are TechCrunch, 18 September 2026, the Wikipedia entry, and DataCamp on System One models and Jev.
Almeida on the gap: "We have lightning in a bottle, and yet it is not useful."
System One, in TypeSafe's naming, means fast focused judgements, not deliberation. The training method is RLCD (Reinforcement Learning for Calibrated Decisions). RLCD is not RLHF, the method Almeida worked on at OpenAI, and it is not RLVR.
Almeida on the data behind that training: "Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data."
Reasoning, coding, and open-ended generation sit outside the class. TypeSafe calls the result a new class of model. The KDnuggets corrective is that TypeSafe optimised an existing shape and did not invent a new task. TypeSafe's own benchmark shows ~68% accuracy. The reference answers come from frontier models, not verified ground truth.
What Jev returns
The input is unstructured program state. The output is typed structured values, plus calibrated probabilities and confidence scores, from that single parallel pass. There is no text output. String generation is given up entirely. The pass returns the typed value instead.
Output cardinality is the count of allowed output values. On Jev it runs up to 255.
The published price and latency
These are TypeSafe's published figures for that pass. This page records them. It does not turn them into a router budget.
| Figure | Published value |
|---|---|
| Input price | $0.042 / MTok ("$42 per billion tokens") |
| Output tokens | Free ("too cheap to meter") |
| End-to-end latency | 70ms to 500ms |
| Versus frontier LLMs | Claimed 40x to 200x faster at comparable intelligence |
| Homepage workflow evals | Claimed 193.6x faster, 444.6x cheaper |
| Output cardinality | Up to 255 |
The frontier-LLM row and the homepage row are separate claims. One is 40x to 200x faster "at comparable intelligence." The other is 193.6x faster and 444.6x cheaper on TypeSafe's own workflow evals. They are not averaged here. Neither row is an accuracy result.
Two hard limits
Jev cannot process images yet. It gives up string generation entirely.
Reasoning, coding, and open-ended generation are outside the class as well. What remains is a judgement over a declared set of up to 255 values, with a probability and a confidence score.
Zero out-of-schema is not zero wrong
Zero hallucinations means zero out-of-schema outputs. It does not mean zero wrong answers. Jev cannot return a value outside its declared schema. It can absolutely still be wrong. This is the single most misread claim about the model.
The line "mathematically cannot hallucinate" is a structural guarantee, not an empirical one. The schema fences the shape of the return. It does not check that the value inside the fence is true.
TypeSafe's own benchmark shows ~68% accuracy. The reference answers come from frontier models, not verified ground truth. Independent evaluation is still thin.
Read the speed and cost claims against that figure. The 40x to 200x range, and the homepage 193.6x and 444.6x, are not a second accuracy result.
KDnuggets states the category correction in three sentences: "Classification is not new. Intent detection is not new. Zero-shot classification is not new." TypeSafe optimised an existing shape. It did not invent a new task.
The probability is a number you read
Bryo AI (Nikhil Mudholkar) uses Jev for business email classification: 10x to 20x cheaper than Gemini, with better confidence scoring.
Armin Ronacher (Earendil CTO) is the caveat on the score. You must interpret the confidence threshold. His line: "if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it." A return inside the schema can still be a value he disregards. The schema did the only thing it guarantees. It kept the value in the set.
On TypeSafe's own terms, a System One model is a fast typed judgement with a calibrated probability and a confidence score, trained with RLCD. It cannot process images yet. It gives up string generation entirely. Output cardinality goes up to 255. Input is $0.042 / MTok, and output tokens are free. TypeSafe's own benchmark shows ~68% accuracy, and the reference answers come from frontier models, not verified ground truth.