Noul yes or no
Start here
Two kinds of AI: one talks, one decides
The AI most people know is a chat model (an LLM): it writes text for a person to read. That is perfect for answering an e-mail. It is awkward when a program needs the answer, because the program has to dig the decision out of a paragraph and hope the paragraph looks the same next time.
Jev skips the paragraph. You describe the possible answers up front, and it returns only those, with a probability for each. TypeSafe calls this a System One model, after the fast, intuitive “System 1” thinking in Daniel Kahneman's book Thinking, Fast and Slow: the kind of snap judgment an expert makes in a few seconds.
Chat model: System 2 style
Writes an answer word by word. Can reason at length, explain itself, write code and letters. Slower and more expensive per decision, and its output is text that software must parse.
Jev: System One
Makes a narrow judgment and stops. Returns typed values and probabilities, so code can branch, sort and route on them. Fast and cheap, but it will not explain itself, reason through many steps or write anything.
Anatomy
One request: a state, some questions, typed answers
Every call has the same shape. The state is what Jev looks at: a message, a document, or a JSON object with several related facts. The questions are what you want to know about it, each with a type and the allowed answers. Back comes one answer per question, under the same name you gave it.
The question name is only for your code
Names like refund or team are keys your program uses to find the answer. TypeSafe's docs say the key is not sent to the model, so the whole meaning must be in the question text and the criteria.
The building blocks
Three question types, three kinds of answer
TypeSafe calls them primitives, like the basic types of a programming language. A Noul is a yes/no question. A Choice picks one option from your list. A Score places the state on an ordered scale you describe. Pick a customer message and watch all three answer it.
Noul
Returns one number, noul, from 0 to 1: the probability that the answer is yes. Around 0.5 means yes and no are about equally likely, not “medium”. No confidence field. Several labels can apply at once? Ask one Noul per label.
Choice
You define up to 255 options, each with a short description. Returns the winning choice, a probabilities map that sums to 1, and a confidence. It is relative: it always picks the best of what you offered, so offer a “none of these” option when nothing may fit.
Score
You describe 2 to 10 ordered levels. Returns the probability of each level and a score: the probability-weighted average position, so 1.57 means “between level 1 and 2, closer to 2”. Good for ranking. Not a ruler: do not read exact quantities out of it.
Knowing when not to act
Probability, confidence, and the three lanes
A Choice or Score answer comes with a full probability distribution. Confidence squeezes its shape into one number: everything on one option gives 1, and the flatter the spread, the lower it goes (for a Choice, an even spread gives 0). That lets the model say “I'm not sure about this one”, and lets your code do something different when it does.
Confidence-gated routing: 14 real tickets
A common pattern from TypeSafe's docs: act automatically when confidence is high, ask for a check in the middle, hand it to a person when it is low. Move the two thresholds and watch the tickets change lanes.
Confidence is not correctness
Look for the “thanks, the new dashboard is great” ticket. No team fits a thank-you, but a Choice must pick one of the options it is given, so Jev sent it to technical, in the English version even with a fairly high confidence (0.83). The fix is not a threshold. It is a better question: we added a fifth option for messages that need no action and asked again.
Speculative fan-out
Ask many questions in one call
Jev reads the state once and answers every question in parallel and in isolation: no question sees another one's answer, so adding questions barely changes the response time. That makes it cheap to ask speculative questions you may not need, and let your code use only the answers that matter.
TypeSafe measured the same effect
In TypeSafe's parallel-questions cookbook, 13 questions about a Wikipedia article on the GDPR came back 12.2 times cheaper and 10 times faster as one call than as 13 calls, with the same answers.
Code stays in control
Small judgments, combined by your code
Do not ask “is this a good startup idea?”. That hides several judgments in one question. Ask one narrow question per factor, then combine the answers with a formula you own. When your priorities change, you change a number in code instead of rewriting a prompt, and nothing needs to be asked again.
What people build
Use cases: anywhere code needs a judgment
A good fit is a decision a knowledgeable person could make in a few seconds, that you need made many times, by a program. Filter by area; each card shows which question types do the work.
Jagged edges
Where Jev goes wrong
TypeSafe publishes a list of known weak spots for each version. For jev-1.13 the short version is: let code do what code can do exactly, and ask Jev only for the judgment in between.
We tested four of them
Each question was asked three times with the same input. Jaggedness means exactly this: some of these it gets right and one it gets wrong, and you cannot tell in advance which.
Numbers from different questions do not add up
Two recorded experiments. Each answer makes sense on its own, but you cannot do arithmetic across questions.
Quiz: code, Jev, or a chat model?
For each task, pick the tool that should do it.
Under the hood
How it is trained, what it costs, how fast it is
Chat models are tuned with RLHF (reinforcement learning from human feedback) to write answers people like. Reasoning models add RLVR (rewards you can verify, such as a correct maths result). TypeSafe trains Jev with what it calls RLCD, reinforcement learning for calibrated decisions: no text at all, just decisions whose probabilities should match how often they turn out right.
RLHF
Human feedback. Made chatbots such as ChatGPT pleasant to talk to. Rewards answers people prefer, which can also reward confident-sounding mistakes.
RLVR
Verifiable rewards. Produced reasoning models that are strong at maths and code, but slower and more expensive per answer.
RLCD
Calibrated decisions. TypeSafe's approach for Jev: return decisions and probabilities instead of text, and make the probabilities honest.
What “calibrated” means
If a well-calibrated model says 0.8 for a hundred different answers, about 80 of them should turn out right. It is a promise about groups of answers, never about one. Calibration is TypeSafe's claim; check it on your own data.
The spec sheet
| Model | jev-latest points to jev-1.13.0 today. Pin the versioned name if you tune thresholds. |
|---|---|
| Price | $0.042 per million input tokens. Output tokens are free. Prepaid credits. |
| Context | 64k tokens per request in total; the state plus the single longest question must fit in 32k. |
| Rate limits | 1,200 requests per minute and 250,000 tokens per second, which TypeSafe says are adjusting while demand grows. |
| Input | Text only: a string, a JSON object or an array. No images, audio or video yet. |
| Languages | English is the primary training language. Others work with lower accuracy; test yours. |
| Customising | No fine-tuning: every account uses the same weights. You shape answers through the state, the question text and the criteria. |
| Data | Processed in the USA (TypeSafe's subprocessors: AWS, Modal, Nebius, CoreWeave), no EU region. Not used for training. Zero data retention is for enterprise customers. |
| SDKs | Python typesafe-sdk, JavaScript @typesafe-ai/sdk, or plain HTTP. |
How fast, really: every call on this page
Each dot is one real call we made to build this guide, timed from our server in Europe to TypeSafe in the USA, network included.
What would it cost?
Who makes it
TypeSafe AI, Inc. is a San Francisco company founded in 2024. It came out of stealth in September 2026 with a 40 million dollar seed round led by DCVC. Its CEO, Diogo Almeida, is one of the authors of InstructGPT, and TypeSafe credits him as a co-inventor of RLHF, the training technique behind ChatGPT. Not to be confused with Typesafe Inc., the unrelated Scala company behind Akka that renamed itself Lightbend in 2016.
Hands on
Ask Jev yourself
Write any customer message and ask the same three questions as above, live. This runs through our AI Classifier API, so it shares that tool's free daily allowance.
The same call from your own code
You need your own key from TypeSafe (console.typesafe.ai). Keep it on your server, never in a web page.
Go deeper: TypeSafe docs · known weak spots of jev-1.13 · every recorded answer on this page (JSON)
Questions
Frequently asked
What is Jev?
Jev is the flagship model of TypeSafe AI, a San Francisco company. It is what TypeSafe calls a System One model: you send it a state (a text, a JSON object or an array) and typed questions about it, and it returns a typed answer with probabilities for each question. It never writes text.
How is Jev different from ChatGPT or another LLM?
A chat model generates text for a person to read, so software has to parse its reply. Jev returns only the answer types you define: a probability of yes, one option from your list, or a position on your ordered scale. That makes it fast and cheap and easy to use from code, but it cannot write a reply, a summary or code.
What are Noul, Choice and Score?
They are Jev's three question types. A Noul is a yes or no question and returns the probability that the answer is yes. A Choice picks one option from a set you define and returns a probability for every option plus a confidence. A Score rates the state on ordered levels you describe and returns the probability-weighted position on that scale, the probability of each level and a confidence.
What is the difference between probability and confidence?
Probabilities say how likely each option or level is. Confidence is one number from 0 to 1 that summarises how concentrated those probabilities are: all of it on one option gives 1, and the more evenly it spreads, the lower it gets (for a Choice, an even spread gives 0). Choice and Score answers carry a confidence; a Noul answer is a single probability and has none. High confidence is not proof that an answer is right.
How fast and how expensive is Jev?
TypeSafe lists the price as 0.042 US dollars per million input tokens, with output tokens free. In our own measurements every call, network included, took between 0.23 and 0.42 seconds from a server in Europe. Asking many questions in one call is much cheaper than one call per question because the state is sent only once.
Does Jev understand Czech and other languages?
English is Jev's primary training language and where TypeSafe says accuracy is best. Other languages are accepted with lower accuracy. The Czech examples on this page were answered sensibly, but they are a handful of texts, not a measurement, so test your own data before relying on it.
What is Jev bad at?
By TypeSafe's own list for version 1.13: reading questions too literally, arithmetic and counting, comparing dates, questions with several levels of indirection, very large states full of irrelevant detail, text written to trick the model, contradictory instructions and criteria, expecting separate answers to add up, and generating text. Keep calculations, date logic and exact lookups in code and ask Jev only for the judgment.
Where is my data processed?
TypeSafe runs Jev in the USA (the subprocessors it lists, AWS, Modal, Nebius and CoreWeave, are all US-based) and has no EU region. It says it does not train on customer requests, and zero data retention is offered to enterprise customers. The recorded examples on this page are invented texts. The live demo sends your text through our AI Classifier to TypeSafe and we do not store it.
In one breath
The whole thing, recapped
- Jev decides, it does not write. A state and typed questions go in; typed answers with probabilities come out.
- Three question types: Noul (probability of yes), Choice (one of your options), Score (a point on your ordered scale).
- Confidence says how concentrated the probabilities are. Use it to choose between acting, checking and asking a person.
- Batch your questions. One call with many questions sends the state once: far cheaper and faster than many calls.
- Keep code in control. Ask narrow questions and combine them with formulas and hard rules you own.
- Mind the jagged edges. Maths, counting and dates belong in code; always offer a “none of these” option; test on your own data.