What is a System One model? Kahneman's fast and slow lanes, in software
By Christophe Paka · Facts checked on 23 October 2026
A System One model is an AI model that makes quick decisions inside your software: your team writes a list of allowed answers in the request, and the model picks one and says how likely each answer is. It never writes text. Jev, from TypeSafe, is the model the name was made for.
So what is a system one model, and why that name? Readers on Reddit ask if it is a new kind of learning, or just a small LLM with a new name. Guess wrong and you give Jev a job that needs a written answer, or pay a slow model for a call that needed one word.
The name comes from a psychology book. The figure below maps that idea onto your software in two rows.
How to read this figure
Read the top left box. The psychologist Daniel Kahneman described two modes of thinking in Thinking, Fast and Slow. System 1 is fast and automatic: answer 2 + 2, read the words on a billboard, see the emotion in a face. You do these with no effort.
Read the top right box. System 2 is slow and takes effort: compute 17 x 24, fill out a tax form, park in a narrow space. You have to stop and pay attention. Most of a day runs on System 1, and System 2 steps in for the hard cases.
Drop to the bottom left: Jev. TypeSafe uses the name for models that make fast decisions in a fixed shape, so code can read them with no text to parse. Example: which team should take this ticket, billing, technical or sales? Jev answers technical, with a probability (how likely each answer is). TypeSafe calls Jev the first model of this kind.
Follow the dashed arrow. When Jev is unsure, your code can escalate: hand the case to someone slower and stronger. The arrow is labeled low confidence. Your code draws that line, not Jev. You pick it by testing on a few hundred cases you already sorted by hand, and a costly action gets a higher line. The post on setting a confidence line for each action shows how.
Read the slow lane and the note. The slow lane is an LLM, a reasoning model or a person. A reasoning model is an LLM that works through steps before it answers, so it is slower and costs more. The note under the figure is a warning: the bottom row is an analogy, not a claim about how Jev works inside.
Key takeaways
- Send quick, repeated decisions with known answers to Jev, and keep slow, open ended work for an LLM or a person.
- Write the escalation rule in your code: below a confidence line you choose, hand the case to the slow lane.
- Expect more uses once each decision is cheap, so budget for more decisions, not just cheaper ones.
Try this: sort one of your decisions into a lane
Pick one decision your code or your team makes many times a day. Write down its possible answers. Then tick each line that is true.
- It repeats at volume: the same kind of judgment about many different messages or records.
- You can list the valid answers in advance: yes or no, a fixed list, or a scale you can describe.
- The input is text, best in English.
- A probability would change what you do: act, ask a person, or hand it on.
- You do not need a written reply or an explanation for this step.
- A young vendor hosting the model for you is acceptable for this step.
Expected result: Five or six ticks: the decision belongs in the fast lane, so try it with Jev. Four ticks: test it on TypeSafe's test page first. Fewer than four: keep it with your LLM.
Questions readers ask
Is a System One model just a tiny, cheap LLM?
No, says TypeSafe: its launch post calls Jev neither small nor an LLM. It does not disclose the model's size or design, so you cannot check that claim. What you can check is the behavior: Jev only picks answers from a list you wrote. Unlike a small trained classifier, it takes the question in each request, so you change a question without retraining.
Why is it called Jev?
Jev is named after William Stanley Jevons. Cheaper steam power raised the demand for coal, and TypeSafe bets cheap judgments go the same way. That is a bet, not a measurement, but plan for it: once each decision costs very little, teams tend to find more decisions to hand over.
What happens when Jev is not sure?
Jev still answers, with probabilities and a confidence number. Your code reads them. If the confidence is below a line you set, your code sends the case to the slow lane: a person or a reasoning model, in TypeSafe's words.
What does one fast lane decision cost?
TypeSafe bills only the text you send in, not the answers. The sample support ticket in its docs cost about $0.000018 to judge, so a thousand like it cost just under 2 cents. The slow lane costs more per case, which is why only the unsure cases should reach it.
Sources: System One models (TypeSafe docs); Introducing System One Models & Jev (TypeSafe).
This figure is from Jev Made Crystal Clear, chapter 1.
The chapter builds the idea step by step, with recipes you can follow, a Try it to check yourself, and every source.
Get the book
