Imagine asking a sharp friend whether the milk has gone off. You do not want an essay about dairy. You want “yes” or “no”, and some sense of how sure they are — because if they say “probably fine?” you are going to smell it yourself.
Almost every AI model in common use answers the essay way. Ask whether an email is urgent and you get a paragraph that means yes. Your code then has to read that paragraph back and work out what it said, which is why so much software around models is spent coaxing them into JSON and checking whether the JSON came out intact.
System One models are an attempt to remove that whole step. They do not write. They take a situation and a set of questions, and return typed answers — a choice, a rating, a probability — with a confidence figure attached. The first one released publicly is Jev, from a company called TypeSafe AI, and it is worth understanding whether or not you ever use it, because the idea behind it is more interesting than the product.
In short
- What
- A model that returns typed decisions instead of text: a choice from a fixed list, a score on a scale, or a probability — each with a confidence number software can branch on.
- Why
- It removes the parse-and-validate step between a model and the code using it, and it gives you a usable signal for when to act, when to review, and when to ask a person.
- When
- High-volume, repeated decisions over a shared state where the possible answers are known in advance — triage, routing, filtering, checking an action before it runs.
- How
- Pair it with a reasoning model rather than replacing one: the language model for open-ended work, the decision model for the fast structured calls along the way.
Where the name comes from
The psychologist Daniel Kahneman described two modes of thinking. System 2 is the slow, deliberate kind — working through a problem, weighing arguments, showing your reasoning. System 1 is the fast kind that happens before you notice it: recognising a face, sensing that a street feels wrong, knowing a sentence is sarcastic.
Most of the effort in AI has gone into making models better at the slow mode, and it has worked remarkably well. TypeSafe’s argument is that a great deal of real automation does not need the slow mode at all. It needs a fast judgement, made the same way ten thousand times an hour, that some piece of software can act on immediately. They call that class of model System One, and describe it as “a new class of frontier models built to make fast, structured decisions that software can use directly”.
Whether or not the framing holds up, the observation underneath it is fair: a lot of what gets built with language models is not conversation. It is classification wearing a conversation’s clothes.
And why it is called Jev
The model’s name is a small joke with a large bet inside it. Jev is short for William Stanley Jevons, a nineteenth-century economist who noticed something awkward about steam engines: as they became more efficient, Britain burned more coal, not less. Cheaper power did not reduce demand for power. It made worthwhile a hundred uses that had not been worth the fuel before.
That is the wager in the name. If a decision becomes two orders of magnitude cheaper, TypeSafe are not expecting anyone’s bill to fall. They are expecting decisions to start being made in places where nobody would previously have bothered — in every list, every queue, every form, continuously. Their founder has put it as “there’s just going to be smart software all over the place”.
It is worth saying plainly that this is a bet, not a result. The Register, covering the launch, pointed out that the assumption the market for tokens will be as broad as the market for energy remains unsettled — plenty of people simply have no use for these tools. Jevons was describing coal after the fact. Nobody has yet described this after the fact.
The company behind it was founded by Diogo Almeida, a former OpenAI researcher who co-invented reinforcement learning from human feedback — the technique that made ChatGPT work as a product — and who has described leaving because the technology, for all its power, was not much use for automation. Knowing that makes the shape of Jev easier to read: it is a deliberate attempt at the opposite of a chatbot, by someone who helped build the chatbot.
What it actually returns
You hand Jev a state — the messages, the record, the policy, whatever the decision depends on — and then ask questions about it. The questions come in three shapes, according to TypeSafe’s documentation.
A Choice picks one option from a list you define in advance, up to 255 of them, and returns the selection along with probabilities for each option. A Score places the state on an ordered scale of between two and ten levels. A Noul returns a probability between zero and one that some statement about the state is true.
Every answer carries a confidence figure derived from the shape of the probability distribution. You ask several questions at once, they are answered in one pass, and what comes back is a structure your code reads directly — no parsing, no repair, no retry because the closing brace went missing.
It does not write text. Not as a limitation to be worked around: it genuinely has no text to give you. That is the trade.
The confidence number is the actually interesting part
Speed is what gets quoted, but calibration is what matters. TypeSafe say the model is trained with something they call Reinforcement Learning for Calibrated Decisions, optimising for “answers with epistemically honest probabilities” — meaning that when it says it is 90% sure, it should be right about nine times in ten.
If that holds, it gives you something ordinary language models are bad at providing: a number you can build a rule on. Above this confidence, act. Below it, put it in a queue for a person. Somewhere in between, escalate to a slower, more capable model that can reason about it properly.
Ask a chat model how confident it is and it will tell you, fluently and with very little relationship to how often it is actually right. Confidence you can trust is a genuinely different tool from confidence you cannot, and it is the part of this worth paying attention to.
“It cannot hallucinate” — true, and not what it sounds like
TypeSafe state that Jev mathematically cannot hallucinate or produce type errors, because the valid outputs are defined in the schema before the question is asked. That claim is correct and worth understanding precisely, because it is easy to hear as something much larger than it is.
What it guarantees is shape. The answer will always be one of the options you listed, always a number in the range you specified, always parseable. It cannot invent a nineteenth category, cannot return prose where you expected a score, cannot produce half a JSON object. Anyone who has written retry logic around a model’s output knows exactly how much tedium that removes.
What it does not guarantee is truth. The Register made this point plainly in its coverage: the comparison “really isn’t a fair comparison as its output is not natural language”, and returning a structured answer “does not preclude the possibility of being incorrect”. A model that must pick one of five categories will always pick one of five categories. It can still pick the wrong one, confidently, all day.
So “zero hallucinations” here means “the output always fits the schema”. That is a real engineering benefit. It is not a claim about accuracy, and reading it as one would be the expensive mistake.
The numbers, and who is holding the stopwatch
TypeSafe publish figures: end-to-end responses in 70 to 500 milliseconds against seconds for frontier chat models, input priced at $0.042 per million tokens with output free, and headline comparisons in the range of forty to two hundred times faster.
The mechanism behind the speed is plausible enough. A language model produces one token at a time, each one waiting on the last, so a long answer is a long wait by construction. A model that returns a fixed structure can answer in a single parallel pass, and answer several questions in that same pass.
The figures themselves should be read for what they are. MarkTechPost noted that the comparisons come from TypeSafe’s own workflow evaluations, written by TypeSafe’s own team, and that the company says it cannot prove the pricing is unsubsidised. That is not an accusation of anything — vendor benchmarks are vendor benchmarks, and early pricing is early pricing. It simply means the honest version of the claim is “fast and cheap on the tasks they chose” until somebody outside has measured it on tasks you care about.
One outside figure has appeared. TechCrunch reported that Vercel, which offers hosted access to the model, measured it at five to eighteen times faster than the model they had been using, with better accuracy on their task. That is a long way below forty to two hundred — and still a large number. It is roughly what one should expect: real gains, smaller than the headline, on somebody else’s workload.
When you would reach for it
The shape of the fit is quite specific: high-volume, repeated decisions over a shared state, where the possible answers are known in advance. Sorting incoming messages by what they are about. Deciding whether a listing needs review before it goes live. Scoring whether a photograph matches its description. Filtering a large set down to the handful worth looking at properly.
The pattern that has emerged fastest in practice is putting one in front of something expensive or risky. LangChain describe using Jev to check an agent’s tool calls and block the dangerous ones before they execute — a cheap, fast judgement standing in front of an action that cannot be undone. That is a harness, in the sense of harness engineering, and a decision model is a very natural thing to build one from.
The other emerging pattern is routing: a small fast model decides which requests are straightforward and which need a slower, stronger model, so that the expensive one is only used where it earns its cost.
When you would not
When the answer is prose. Writing, summarising, explaining, drafting a reply — none of that is available here, and no amount of clever schema design gets around it.
When you cannot list the possible answers in advance. Open-ended work, genuinely novel situations, anything where the useful response is a category nobody anticipated: a fixed set of options is the wrong instrument.
When you need the reason as well as the verdict. A probability is not an explanation. If a person is going to be told why something was rejected, something still has to write that sentence.
And when the volume simply is not there. A decision made a few hundred times a day does not need to be two hundred times faster, and adding a second model to your stack has a cost that is not measured in tokens.
What to keep an eye on
At the time of writing this is early-access software behind a waitlist, available as a hosted API. TypeSafe have not published weights, a parameter count or a self-hosting option, which means anything built on it is built on a dependency you do not control and cannot run yourself.
Calibration is the claim to test, not take. It is also the one you can check yourself, and cheaply: collect a few hundred decisions with their confidence figures, wait until you know the real answers, and see whether the nineties really are right nine times in ten on your data rather than on the examples in the announcement.
And the broader idea is worth separating from the specific product. Whether or not Jev is the model that lasts, “return a typed decision with honest uncertainty instead of a paragraph” is a good idea that software has wanted for a while. If it works, others will build it.
Where it stops
A decision model decides. It does not explain itself, it does not write, and it knows nothing except what is in the state you handed it — so the quality of its answers is bounded by the quality of that state, which is context engineering wearing a different hat.
It is a component, not a replacement. The useful mental model is not “this instead of a language model” but “this for the thousand small judgements, that for the handful of hard ones” — which is roughly how people work too.
