Jev and System One models: a plain guide / Start here
What is a System One model?
A System One model is a model that answers questions you wrote in advance and returns a probability for each possible answer, without generating any text. TypeSafe AI launched the first public one, Jev, on 15 September 2026 and opened access to everyone on 20 September. It fits tasks such as classifying, routing, scoring, extracting a label and choosing a branch. It cannot chat, write code or explain a decision. The output shape is what makes it faster and cheaper than a language model for the decisions it fits.
Published September 22, 2026. Editorial.
Key takeaways
- A System One model returns a decision with a probability per answer: the possible answers are fixed before the request is sent and the model returns a probability for each.
- The name borrows Daniel Kahneman's split between fast automatic judgment and slow deliberate reasoning, and the model does only the fast kind.
- Tasks that fit are classify, route, score, extract a label and branch. Tasks that do not fit are chat, code, summaries and any answer that has to be explained in words.
- Because the answer is a set of probabilities over answers you defined, there is no output to generate, which is why TypeSafe can price output at zero and report response times under half a second.
- Reveneau uses Jev for exactly this class of task inside its eval suite: the fixed-answer checks that used to need a language model judge.
On 15 September 2026, TypeSafe AI published a post introducing Jev and the category it belongs to. The definition it gave was "a new class of frontier models built to make fast, structured decisions that software can use directly" [1]. Five days later the waitlist came off and anyone could create a key at console.typesafe.ai [2]. We had a use for it the same week, and this page is the definition we wish we had read first.
The definition, in plain words
A System One model takes two inputs. The first is a piece of text, which TypeSafe calls the state: a string, a JSON object, or an array of text values, and text only, with no images or audio. The second is a set of named questions. For each question, the model returns a typed answer with probabilities [3]. A question might ask which of five teams should handle a support ticket, how severe a bug report is on a scale of 0 to 3, or whether a paragraph contains medical advice. The answer comes back as a probability per option, a probability-weighted position on the scale, or a single probability that the statement is true.
What the model never returns is a sentence. There is no reply to read, no explanation, and no generated text of any kind. That is the whole difference from a language model, and everything else on this page follows from it.
Why the name
Daniel Kahneman described human thinking as two systems. System One is fast, automatic, and effortless: recognising a face, sensing that a sentence is hostile, sorting mail into piles. System Two is slow, deliberate, and effortful: working through a proof, writing an argument, planning a trip. TypeSafe borrowed the first label for a model that does the first kind of thinking only.
The label is more useful than it looks, because it gives you a test you can apply to any task in your product. Ask whether a competent person would answer it in a second without writing anything down. Which queue does this go to. Is this message spam. Does this response follow the instruction. Those are System One tasks. Ask whether the person would need to think, draft, and revise. Write the reply. Fix the bug. Explain why the decision is right. Those are System Two tasks, and they need a language model. We wrote a longer version of that test in how to tell a System One task from a System Two task.
What kinds of task fit
The tasks that fit share one property: the set of possible answers is known before the request is made. Five kinds cover most of what a product needs.
Classify. Put a text into one of a fixed set of categories. Support tickets by topic, reviews by sentiment, documents by type. This is the oldest task in machine learning and Jev's Choice question does it directly, with up to 255 options per question [4].
Route. Decide which system, model, or person should handle a request. TypeSafe's own patterns page names intent routing as one of its four core patterns, and LangChain's integration uses a TypeSafe classifier to pick which language model should answer a request [5]. Routing is a classification whose output is a branch in your code.
Score. Place a text on an ordered scale. How severe is this report, how complete is this answer, how far did this response drift from the instruction. Jev's Score question returns a probability-weighted value on a scale of 2 to 10 levels you describe in words [6].
Extract. Pull a label out of text. Which product does this complaint mention, which of these three dates is the deadline, which clause is the one about termination. Extraction of a value from a fixed set is a Choice question. Extraction of free text, such as a name nobody listed, is a language model task.
Branch. Decide yes or no. Is this a jailbreak attempt, does this diff touch the payment path, did the agent complete the goal. Jev's Noul question returns one probability that the statement is true [7].
What does not fit
Anything that needs a sentence. A chat reply, a code change, a summary, a rewrite, a translation. The model is not trained to generate text and TypeSafe lists that on its own page of known weaknesses [8].
Anything that needs an explanation. If your product has to tell a user why a decision was made, the decision model cannot supply the reason. You can log the probabilities, and they are useful for tuning, but a probability of 0.83 measures the model's belief, and a person who wants a reason needs a sentence from a language model. A common pattern is to let Jev decide and then, only on the branches that need it, ask a language model to write the explanation for the decision already made.
Anything whose answers are open-ended. If the right option might be one nobody put on the list, a fixed list is the wrong tool. The model will pick the closest option with whatever probability it has, and nothing in the output tells you that the true answer was missing.
Anything numeric. TypeSafe's jaggedness page says the model does not count reliably, cannot reliably judge whether two numbers are near each other, and reads dates as text [8]. Those belong in code, and where Jev is weak goes through each one.
Why returning a decision changes cost and speed
A language model produces its answer one token at a time, and every token is a full pass through the model. Ask it to classify a ticket and it will write "This ticket concerns a billing dispute and should be routed to the finance team" and then, if you asked nicely, a JSON object you can parse. Each of those tokens costs money and time, and the JSON sometimes comes back in the wrong shape.
A System One model has no tokens to generate. The answer is a set of probabilities read out in one pass. That is why TypeSafe can price output at zero and charge only for input, at $0.042 per million input tokens [9], and why it reports response times of 70 to 500 milliseconds end to end against 3 to 329 seconds for the frontier language models in its own comparison [1]. TypeSafe's headline benchmark, reported by MarkTechPost, has Jev finishing a workflow decision in 0.114 seconds for $0.000081 against GPT-5.6 Terra at 8.566 seconds for $0.013880 [10]. Those workflows were written by TypeSafe's own team, so treat the ratio as the vendor's best case. The direction of the difference does not depend on the workload, though. No output means no output cost and no output latency.
The same property removes a whole class of failure. The docs describe the answer as always matching the schema you asked for, which TypeSafe calls not hallucinating. What that means in practice is 0 percent structured-output errors on TypeSafe's evaluation, against 45.5 percent for Claude Haiku 4.5 on the same workflows, as DataCamp reports [11]. The answer can still be wrong. On that same evaluation Jev agreed with the reference answer 67.8 percent of the time [11]. A well-formed answer and a correct answer are different things, and the probability that comes with every answer is how you tell them apart. Calibrated probabilities and confidence, explained covers how to use it.
How it fits next to a language model
The right mental model is a division of labour, and TypeSafe's own documentation for coding agents says so: Jev "is not a drop-in replacement" for the language model running an agent. It is used inside the application for routing, rubric scoring, and statement verification [12]. The language model does the System Two work of writing code and replies. The decision model does the System One work of deciding what to do with what was written.
We use it exactly that way. Reveneau generates all of its code with AI and gates every change with a large eval suite written from the specification. Checks that have a deterministic answer stay deterministic. The checks that need a judgment with a fixed answer, such as whether a change matches an acceptance criterion as written or whether a diff stays in scope, are now Noul or Score questions to Jev, and on our own suite the run is ten times faster than it was with a language model grading. How Reveneau uses Jev describes the setup and evals with Jev is the how-to for your own suite.
The launch, briefly
TypeSafe announced Jev on 15 September 2026 and named its training method Reinforcement Learning for Calibrated Decisions, or RLCD. It has not disclosed the architecture, as MarkTechPost noted on 19 September [10]. The current model is jev-1.13.0, released the same day as the announcement, with the aliases jev-latest and jev-preview both pointing to it [9]. The waitlist was removed on 20 September, according to the unaffiliated systemonemodels.org hub [2]. Three teams published what they built in the first week: LangChain tested it as an agent evaluator, Openlayer released the jevals library, and Browser-use built a browser agent on it. The pillar page, Jev and System One models, summarises what each of them found.
A test you can run tomorrow
Pick one decision in your product that a language model currently makes and whose answers are a fixed list. Write it as one Jev question, send it the same inputs the language model gets, and compare the answers against the language model's on a few hundred cases. You will learn three things: whether the two agree, how far apart the response times are, and what the probabilities look like on the cases where they disagree. That last one is the most useful, because it tells you where to set the threshold. Jev vs a language model has the decision rule for what to move and what to leave.
A System One model is a narrow tool. The narrowness is what makes it fast, cheap and predictable, and it is what makes it wrong for anything that needs a sentence. Know which of your decisions are which and the choice makes itself.
Best for
- Decisions whose possible answers are a fixed list you can write down before the request
- Decisions that repeat at volume, where seconds and cents per call add up
- The judgment step inside a pipeline where code already handles the arithmetic and the side effects
Avoid if
- Do not use a decision model for any output a person will read as a sentence
- Do not use it when the right answer might be an option nobody put on the list
- Do not send it a counting, comparison, or date question, because TypeSafe documents all three as unreliable
Check before you decide
- Confirm that each question has a fixed set of answers and that your code can act on every one of them
- Confirm agreement against your current method on a labelled sample before switching any live traffic
- Confirm what the probabilities look like on the disagreements, since that is where the threshold goes
Common questions
What is a System One model?
A System One model is a model that answers questions written in advance and returns a probability for each possible answer, with no generated text. TypeSafe AI launched the first public one, Jev, on 15 September 2026. The test for whether a task fits is whether a competent person would answer it in a second without writing anything down: routing, classifying, scoring and yes/no checks fit, while replies, code and explanations do not.
Why is it called System One?
The name comes from Daniel Kahneman's split between two kinds of thinking: System One is fast, automatic judgment, and System Two is slow, deliberate reasoning. TypeSafe took the first label for a model that does the fast kind only. The practical use of the name is as a sorting rule: if the task needs drafting or revising, it is System Two and needs a language model; if it needs an instant call from a fixed list, it is System One.
What tasks fit a System One model?
Five kinds cover most products: classify a text into one of a fixed set of categories, route a request to a system or person, score a text on an ordered scale, extract a label from a fixed set, and branch on a yes/no statement. Each maps to one of Jev's three question types. The shared property is that every possible answer is known before the request is sent, so code can act on any of them.
What tasks do not fit?
Anything that needs a sentence, an explanation, an open-ended answer, or arithmetic. Jev is not trained to generate text, so replies, code, summaries and rewrites are out. It cannot supply a reason for a decision, only a probability. If the right answer might be an option nobody listed, a fixed list is the wrong tool. And TypeSafe documents counting, numeric comparison and date ordering as unreliable, so those belong in code.
Why does returning a decision instead of text make it cheaper?
A language model writes its answer one token at a time and each token costs money and time. A System One model has nothing to write: the probabilities are read out in one pass. That is why TypeSafe charges $0.042 per million input tokens and nothing for output, and why it reports 70 to 500 milliseconds end to end. The ratio depends on the workload, and the direction of the difference does not.
Does a System One model hallucinate?
TypeSafe says Jev cannot hallucinate, and what it means is that the answer always fits the schema you asked for: 0 percent structured-output errors on its own evaluation, against 45.5 percent for Claude Haiku 4.5, as DataCamp reports. The answer can still be wrong. On the same evaluation Jev agreed with the reference 67.8 percent of the time, so the probability attached to each answer, and a threshold on it, are still required.
When did Jev become available?
TypeSafe announced Jev on 15 September 2026 with a waitlist, and the current model jev-1.13.0 was released the same day. The waitlist was removed on 20 September 2026, according to the unaffiliated systemonemodels.org hub, and anyone can now create a key at console.typesafe.ai. Within that first week LangChain, Openlayer and Browser-use each published something built on it.
How does a System One model work next to a language model?
As a division of labour. TypeSafe's own coding-agents documentation says Jev is not a drop-in replacement for the language model running an agent and is used inside the application for routing, rubric scoring and statement verification. The language model writes the code or the reply. The decision model decides what to do with what was written, in under half a second, and its probability gates whether the result goes out automatically or to a person.
Is a System One model the same as a classifier?
It does the same job as a classifier for a Choice question, with one difference in how you use it: the categories are described in words in the request rather than learned from labelled data, so you can change them without retraining. A classifier you train yourself is still the better choice when you have a large labelled set, when data cannot leave your network, when the text is mostly outside English, or when the decision is numeric.
What is the fastest way to find out whether a decision fits?
Take one decision a language model currently makes whose answers are a fixed list. Write it as one Jev question, send it the same inputs, and compare the two answers on a few hundred labelled cases. You learn whether they agree, how far apart the response times are, and what Jev's probabilities look like on the disagreements. The disagreements are where the threshold goes, so they are the most useful part of the test.
How does Reveneau use System One models?
Reveneau writes all of its code with AI and gates every change with an eval suite written from the specification. Checks with a deterministic answer never touch a model. The checks that used to need a language model judge, such as whether a change matches an acceptance criterion or whether an agent's trace followed the plan, are now Noul or Score questions to Jev. On our own suite the run is ten times faster than with the previous grader.
References
- [1] TypeSafe AI, Introducing System One models and Jev (15 September 2026): "a new class of frontier models built to make fast, structured decisions that software can use directly"; 70 to 500 ms end to end against 3 to 329 seconds; "40x-200x faster"; RLCD named as the training method.
- [2] systemonemodels.org, Jev model page (unaffiliated with TypeSafe): waitlist removed on 20 September 2026, sign-up open at console.typesafe.ai.
- [3] TypeSafe docs, System One concept: state is a string, JSON object, or array of text values, text only; questions are a map of named questions; output is a typed answer per question with probabilities; no text generation.
- [4] TypeSafe docs, Choice primitive: pick one of up to 255 options; returns the choice, a probability per option, and a confidence number.
- [5] LangChain, Building a harness with Jev: the langchain-typesafe package and TypeSafeClassifier used for model routing and a middleware that blocks risky tool calls before execution.
- [6] TypeSafe docs, Score primitive: an ordered spectrum of 2 to 10 levels described in words; the score is each level multiplied by its probability, added up; example score 1.43, confidence 0.35, probabilities {0: 0.0, 1: 0.57, 2: 0.43}.
- [7] TypeSafe docs, Noul primitive: a yes/no statement; returns one probability from 0 to 1; no separate confidence number.
- [8] TypeSafe docs, Model jaggedness for jev-1.13: literal reading, unreliable counting, unreliable numeric comparison, dates read as text, and not trained to generate text.
- [9] TypeSafe docs, Models: $0.042 per million input tokens, output free; jev-1.13.0 released 15 September 2026; aliases jev-latest and jev-preview.
- [10] MarkTechPost, TypeSafe AI releases Jev (19 September 2026): 0.114 s and $0.000081 against GPT-5.6 Terra 8.566 s and $0.013880; workflows written by TypeSafe's own capabilities team; architecture undisclosed.
- [11] DataCamp, Jev: TypeSafe's System One model explained: 0% structured-output errors for Jev against 45.5% for Claude Haiku 4.5; 67.8% agreement for Jev on TypeSafe's own workflow evaluation.
- [12] TypeSafe docs, Coding agents: Jev "is not a drop-in replacement" for the LLM running a coding agent; used for routing, rubric scoring, and statement verification.
Related reading
How to tell a System One task from a System Two task
If the possible answers can be written down before you ask, the task repeats at volume, and nobody needs a written reason, a fast decision model can do it. Everything else needs a language model or a person. Here is the rule, with three features walked through it.
Our eval suite now runs ten times faster. Here is what we changed.
This month we replaced the language model that graded the judged checks in our eval suite with Jev. The deterministic checks did not move, every rubric is now written inside the check, and the suite runs ten times faster.
AI evaluation and guardrails for production: how to know your AI actually works
"It seems to work" is not a production standard. Here is how we turn that feeling into a number, and how we stop the system from doing damage on the days the number is bad.
Can you trust an AI agent with real work yet?
An agent that answers a question and an agent that takes an action are not the same risk. Here is how we decide where an agent is ready to act, and where it is not.