Jev returns decisions instead of text, at $0.042 per million tokens. Where TypeSafe's new model fits in an agent stack

On Monday 15 September, a San Francisco lab called TypeSafe AI came out of two years in stealth with a model named Jev. The launch post opens with a claim that most model launches this year would not dare make: Jev does not generate text. You send it a piece of text and a set of typed questions, and it returns answers, probabilities and a confidence score in a single pass. There is no reply to parse, no JSON to validate, and no risk of the model wandering off into an explanation nobody asked for.

The reaction was immediate. TechCrunch ran it as a new kind of model thrilling developers. The Register noted, fairly, that a model which cannot write prose cannot hallucinate prose either, which is a weaker claim than "hallucination-free" sounds. LangChain published a guide to wiring it into an agent within four days. This piece is for the people who run agents in production and want to know whether Jev belongs in their stack, and where.

What is Jev, and what does a System One model do?

Jev is what TypeSafe calls a System One model: a model trained to make fast, structured judgements that software can consume directly, rather than to write text for a person to read. The name borrows Daniel Kahneman's split between fast automatic thinking and slow deliberate reasoning. TypeSafe's version is narrower than the metaphor. The model evaluates a state, which is any text up to about 32,000 tokens, against questions you define, and returns one of three answer types.

A noul question is a yes-or-no judgement returned as a probability: "Does this message express urgency?" comes back as 0.999. A choice question picks one option from a list you supply, up to 255 of them, and returns the probability of every option alongside the pick. A score question places the state on a scale between labelled levels and returns a continuous value, so a support ticket can score 1.035 on a scale where 1 means "frustrated but civil" and 2 means "very angry", which tells you it sits just past civil. Every answer carries a confidence figure, which TypeSafe says is calibrated by a training method it calls reinforcement learning for calibrated decisions.

The quickstart example is a support ticket. One call asks which department should handle it, how frustrated the customer is, and whether it is urgent, and gets all three answers back together. That is the shape of most decisions inside an agent workflow: classify, route, score, extract a flag, decide whether to branch. Today those decisions are made by asking a frontier model for JSON and hoping the parse succeeds.

What does Jev cost compared with a frontier LLM?

Jev charges $0.042 per million input tokens and nothing for output. TypeSafe's own comparison table puts frontier LLM input pricing at $0.20 to $10 per million tokens, with output about five times more expensive again. Claude Fable 5.1 and GPT-6 Astra both list at $10 in and $50 out, as we set out earlier this month. On a 300-token ticket with a 50-token JSON answer, a frontier call costs roughly $0.0055 and Jev costs about $0.000013. TypeSafe's headline figure of 444.6 times cheaper comes from its own workflow benchmark and the company says real-world gains will sit below that. Even at a tenth of the claim, a routing step that ran a thousand times a day for $5.50 would run for pennies.

Speed matters as much as price for this class of work. TypeSafe quotes 70 to 500 milliseconds per call, against 3 to 329 seconds for frontier models on the same decision. The Doom demo on their site makes ten calls a second and the engineer who built it worked out the bill at about $7 an hour. Nobody will run a business process at ten decisions a second, but a document pipeline that scores every paragraph of every contract now can.

The published limits for the current model, jev-1.13.0, are 64,000 tokens of context per request, of which 32,000 is available for the state plus the longest question, 250,000 tokens per second, and 1,200 requests per minute. Input is text only: strings, JSON objects or arrays of text. Images, audio and video are listed as unsupported for now. Access is early access, through a single endpoint and a Python SDK, and the pricing page carries a line I have not seen a vendor write before: the company cannot prove its pricing is not subsidised, and expects it to fall rather than rise.

How accurate is Jev, and what do the benchmarks leave out?

TypeSafe reports about 67.8 per cent accuracy on an internal benchmark of four production workflows, which it says is comparable to a mid-tier frontier model such as GPT-5.6 Terra. Every, which ran an independent test, found Jev roughly 25 times faster and 580 times cheaper than Claude Fable 5.1 on an extraction task, which supports the direction of the claims if not the magnitude.

Three caveats, all of which TypeSafe discloses or which reporters have already raised. First, the workflow benchmark measures agreement with GPT-6 Astra and Claude Fable 5.1 rather than agreement with a human-labelled ground truth, so 67.8 per cent means "agrees with the big models two times in three", which is a different number from "is right two times in three". Second, Every's test covered one extraction task; nobody outside the company has published results across a range of workloads. Third, TypeSafe has no named production customers and no disclosed revenue. The $40 million seed round and reported $200 million valuation tell you investors like the bet, and nothing more.

The most useful page on TypeSafe's site is the one titled model jaggedness, reviewed on 17 September, which lists what Jev 1.13 is bad at. Literal reading. Arithmetic and numbers. Comparing dates and times. Indirection, where the answer depends on following a reference. Large states padded with irrelevant detail. Adversarial content. Contradictory criteria. The company's advice for the arithmetic case is to keep the arithmetic in code, which is the right advice for any model and rarely stated so plainly.

Where does Jev fit in an agent architecture?

Jev belongs at the decision points inside a workflow, and not at the points where a workflow has to produce something a person or another system will read. In our March piece on agents versus automation we argued that most business processes are pipelines with a handful of fuzzy branches, and that the fuzzy branches are where a model earns its keep. Jev is built for exactly those branches.

Concretely, the steps that fit are triage and routing of tickets, emails and documents; scoring for priority, sentiment, risk or completeness; extraction of flags and categories from unstructured text; the guardrail that checks a frontier model's output before it is sent, or checks a user's input before it reaches a tool; and map-reduce style feature extraction across large document sets, where the cost of a frontier call per document has made the job uneconomic until now. TypeSafe lists judging, verifying and jailbreak detection of LLM prompts and outputs among the intended uses, and a cheap calibrated judge sitting in front of an expensive agent is a pattern we already recommend.

The steps that do not fit are anything that has to write: a customer reply, a summary, a piece of code, a negotiation turn. Anything that depends on arithmetic or date logic, which belongs in code regardless. And anything where the input is hostile by design, since adversarial content is on the jaggedness list and a routing model that can be steered by the text it routes is a security problem, as we wrote about agent permissions in August.

The calibrated probabilities change how you design the branch. With a frontier model returning a label, you get "billing" and have to trust it. With Jev you get billing at 0.84, technical at 0.159 and a confidence of 0.596, and you can write the rule: auto-route above 0.9, ask a person between 0.6 and 0.9, and escalate below that. That is the human-in-the-loop threshold most teams want and few can implement, because most models do not give them a number to threshold on.

How should a company trial Jev without betting a workflow on it?

Run it in shadow behind whatever makes the decision today, and compare. Pick one decision your pipeline already makes with a frontier model or a rule set, send the same inputs to Jev in parallel, log both answers, and measure agreement against a human-labelled sample of a few hundred cases. This is the same evaluation discipline we described for evaluation harnesses in April, and it costs almost nothing at Jev's prices. Two weeks of shadow traffic will tell you more than any benchmark, including TypeSafe's.

Then treat it as a vendor with the risk profile of an early-access startup. Keep the frontier path wired in as a fallback and make the switch a configuration change. Record the model version you qualified against, because jev-1.13.0 will not be the last. Do not send it personal data you have not agreed to send to a new processor; the privacy terms are a week old and your data protection officer has not read them. And keep the budgeting maths honest: metered agent workloads get cheaper with a model like this, but a decision step that becomes 400 times cheaper tends to be run 400 times more often.

If the shadow numbers hold, the saving on the one decision will be the smaller part of the payoff. The larger part is the decisions you were skipping because a frontier call per item was too slow or too expensive, and can now afford to make.

Frequently asked

What is Jev by TypeSafe AI? Jev is a model released by TypeSafe AI on 15 September 2026 that reads text and returns typed decisions rather than generated text. You send it a state, which is any text up to about 32,000 tokens, and questions of three types: noul (a yes-or-no probability), choice (one option from up to 255, with a probability for each) and score (a continuous value between labelled levels). Every answer includes a calibrated confidence figure. TypeSafe calls this a System One model and says it is the first of a new class.

How much does Jev cost and how fast is it? Jev is priced at $0.042 per million input tokens, with output tokens free. TypeSafe quotes response times of 70 to 500 milliseconds per call, compared with 3 to 329 seconds for frontier LLMs, and claims up to 193.6 times faster and 444.6 times cheaper on its internal workflow benchmark. An independent test by Every on one extraction task found roughly 25 times faster and 580 times cheaper than Claude Fable 5.1. Rate limits are 250,000 tokens per second and 1,200 requests per minute, and the model is in early access.

What is Jev not good for? Jev cannot write text, code or explanations, so it does not replace a chat or drafting model. TypeSafe's own jaggedness page lists literal reading, arithmetic, date and time comparison, indirection, large inputs full of irrelevant detail, adversarial content and contradictory criteria as known weaknesses. Its accuracy benchmark measures agreement with frontier models rather than human ground truth, and the company has no named production customers yet, so it should be trialled in shadow before any workflow depends on it.

Related reading

• GPT-6 Astra, Claude Fable 5.1 and a $12.9 billion Hugging Face deal in one week
• AI agents vs automation: when to use which
• Building evaluation harnesses for production AI systems
• Agent permissions: how to decide what an AI agent may read, write and spend
• Flat-rate AI is ending for agents. How to budget when every loop is metered

A model that gives you a probability instead of a paragraph changes how you design the routing, scoring and guardrail steps of an agent, and those steps are most of what an agent does. Deciding which of your decisions belong on a model like this, and proving it with shadow traffic before you switch, is the kind of work we do in an AI agents engagement.