What Jev is, and why it never writes a sentence
Jev, TypeSafe AI's first System One model, returns typed decisions with confidence scores, not text. Its claims, its place beside an LLM and open questions.
Most AI tools you have used write text. You ask a question and get back a paragraph, a list or a block of code. Jev, from TypeSafe AI, does not write a sentence. It answers a set of small, typed questions about a piece of information, such as “which team should handle this email?”, and returns a result that software can act on directly.
That matters for a business because many everyday decisions are small and repetitive: sort this message, flag this invoice, rank this lead. TypeSafe claims Jev makes them in well under a second for a fraction of a language model’s cost. Below we set out what Jev returns, what the company claims and cautions, where it fits next to a language model, and what is still unknown.
Fast decisions, in the System 1 sense
TypeSafe calls Jev its flagship model and the first of a class it names System One models, “built to make fast, structured decisions that software can use directly.” The name draws on Daniel Kahneman’s distinction between System 1 and System 2 thinking, from Thinking, Fast and Slow: System 1 is quick and intuitive, System 2 slow and deliberate. Jev itself is named after William Stanley Jevons.
The idea is that many business decisions belong to the first group. Is this email a complaint? Which team should this ticket go to? Does this invoice match its purchase order? A person answers these in a moment, and software could too, if the answer came back in a predictable shape.
What Jev returns instead of text
Jev takes a piece of state, either text or an object, plus a set of questions. Each question has one of three types:
- Choice picks one option from a list you supply.
- Score rates the state against a rubric you write.
- Noul rates a yes-or-no style statement with a value from 0 to 1.
Choice and score answers come with probabilities and a confidence value. The company says the possible outputs are defined in advance and that “the model never makes type errors,” a guarantee it describes as true by construction rather than measured. That is a promise about the shape of the answer, not its truth: you will always get one of your options back, but it can still be the wrong one.
Because the output is a structure rather than prose, there is nothing to parse. Your software reads a field and branches on it. The trade-off is that Jev gives no reasoning. You see the decision and the confidence, not why.
TypeSafe also says all answers carry calibrated probabilities and confidence scores. Calibration means that when a model reports 90 percent confidence, it should be right about nine times in ten. That is a testable claim, and the most useful one to test.
What TypeSafe claims, and what it cautions
TypeSafe reports an end-to-end response time of 70 to 500 milliseconds, and says that is 40 to 200 times faster than frontier language models (the largest general-purpose models) on System One-shaped queries. On price, it lists $0.042 per million input tokens ($42 per billion), with output tokens free, against frontier model input prices it puts at $0.20 to $10 per million. These are the company’s figures at the time of writing (late October 2026).
Its workflow evaluation covers four workflows: security incidents, agent trace observability, invoice processing and customer service. There, Jev scored 67.8 percent accuracy at a mean of 0.4 seconds and $0.0004 per case. That is below the strongest language models in the company’s own table, but at a small fraction of their cost and time.
To its credit, TypeSafe states the limits plainly. It says its reference answers bias results toward OpenAI and Anthropic models, that its own team wrote the workflows, that speed tests were run from laptops on the West Coast, and that it “can’t prove” its pricing isn’t subsidised. Treat them as a reason to run your own test, not as a result.
Where Jev fits next to an LLM
The sensible way to think about Jev is as a decision layer beside a language model, not a replacement for one. Jev decides; an LLM writes.
Take a property management company’s inbox. A fast typed check routes each message to maintenance, leasing or accounts, and scores how urgent it looks. A burst pipe goes to the top of the queue. A language model then drafts a reply for a person to review. Anything the fast check is unsure about goes to a person, or to a larger model. DataCamp’s review of Jev recommends the same pattern: a fast decision and guardrail layer, with harder cases escalated to an LLM.
The model is rarely the hardest part of a project like this. The harder questions are which decisions are worth making in milliseconds, and what happens when the answer is wrong. That is where we start when we design AI agents for a business workflow, and it is the same thinking as a process audit before automation.
What is still unknown
TypeSafe does not say whether Jev’s weights are open, meaning whether the trained model files can be downloaded and run by others, and it does not state a licence for the model. For now, assume a hosted API. Our post on open weight versus open source explains why the distinction matters.
The architecture is undisclosed beyond three names: a new model architecture, a parallel sampler, and a training method the company calls Reinforcement Learning for Calibrated Decisions. TypeSafe has not described that method in detail.
Access is limited too. Jev is in early access, and the company has published no pricing tiers or rate limits alongside its per-token prices. When DataCamp reviewed Jev in September, no independent reproduction of the benchmark had surfaced.
So the first step is a small one. Pick one decision a person makes many times a week, gather a few hundred past cases with known answers, and see how often Jev agrees, and whether its confidence means what it says.
Sources
Start with the people doing the work.
We build software, AI chatbots, and AI agents around the systems you already run, with people keeping the judgment.