TypeSafe AI shipped Jev on 15 September 2026
The model that won’t talk to you — and why that’s the point.
Jev is a System One model: unstructured state in, typed probabilistic decisions out. No strings, no parsing, no retry loop around json.loads. These are independent, hands-on tutorials for it — written the week it launched, with every claim traced to a source.
One request
{
"state": "Our API started returning 500s 20 minutes ago
and we can't process orders until it's fixed.",
"model": "jev-latest",
"questions": {
"department": {
"type": "choice",
"instructions": "Which team should handle this",
"criteria": {
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions"
}
},
"frustration": {
"type": "score",
"instructions": "How frustrated the customer appears",
"criteria": ["Calm", "Frustrated but civil", "Very angry"]
},
"is_urgent": {
"type": "noul",
"instructions": "The message conveys urgency"
}
}
}Three typed answers, evaluated in parallel
- technical0.85
- billing0.08
- sales0.07
- Very angry0.65
- Frustrated but civil0.30
- Calm0.05
no ← 0.5 → yes · nouls carry no confidence field
No prose. No JSON inside a string. Nothing to parse, nothing to validate, and no code path for “the model replied with an apology instead”.
$0.042/ MTok in
output not charged
70–500ms
TypeSafe's figure
255options
max Choice cardinality
64ktokens
32k state + longest question
System Two vs System One
It gives up strings. In exchange it gets four things back.
Three of them are speed and cost arguments. One of them — calibration — is the thing JSON mode on your existing model genuinely cannot give you.
Condensed from TypeSafe’s launch post. The last row is ours — it is the constraint most likely to rule Jev out for you.
Before you commit
Three things the launch coverage got wrong
We like this product. We also read the footnotes. Every number below is TypeSafe’s own — the corrections come from their documentation, not from a competitor.
- 01
“Never hallucinates”
It never makes a type error, which is structural and genuinely useful. It can absolutely return the wrong valid answer. The 0% bar on TypeSafe's own chart is, in their words, “not empirical” — it is added by definition.
Read the full argument → - 02
“193.6× faster, 444.6× cheaper”
One best-case demo, which TypeSafe itself flags as “on the higher end”. Matched on accuracy against its nearest peer it is roughly 76× cheaper and 25× faster — and 6.3 points behind the best baseline. Independent testers measured 3–11×, and on one many-class task Jev simply lost.
Read the full argument → - 03
“Calibrated probabilities”
TypeSafe published no ECE, Brier score or reliability diagram at all. A pre-registered community study since has: calibrated on one dataset (ECE 0.020), overconfident on another (0.094). Calibration is dataset-dependent, so the only number that matters is the one from your data.
Read the full argument →
Latest
It has already changed since launch
Tutorials
Start at the first call. Finish in production.
Your first Jev call
Get a key, send one request, and read a typed answer back — in cURL, Python and TypeScript.
ReadNoul, Choice and Score
Jev has exactly three question types. Picking the wrong one is the most common early mistake — here is how to tell them apart.
ReadDesigning state Jev can actually use
State is not a prompt. Shape it like a record you'd hand an expert, filter it in code first, and name the parts your questions point at.
ReadAsk every question at once
Questions run in parallel and output tokens are free, so asking something you might not need is close to free. Branch in code afterwards.
ReadConfidence-gated routing
The answer tells you what. Confidence tells you whether you're allowed to act on it without asking anyone.
ReadComposite scoring: move the weights into code
Stop asking one model one big question. Ask several small ones and own the weighting yourself.
ReadConcepts
What is actually new here
What a System One model is
A new model class, or a classifier with better marketing? The honest answer is somewhere in between, and the distinction that matters is narrower than either camp says.
ReadCalibration, honestly
The central claim of the whole product, stated precisely — and what independent testers found when they finally measured it.
ReadWhat "cannot hallucinate" actually guarantees
A real, structural, useful guarantee — about the type of the answer, and nothing whatsoever about its correctness.
ReadOne request. A dozen typed decisions. No parsing.
Whether or not Jev is the model you end up shipping, the architecture it forces on you — decomposed judgments, thresholds in code, policy in version control — is worth adopting today.