Jev by typesafe.ai
TypeSafe AI: Introducing System One Models & Jev
What is it?
A model that doesn't generate text. TypeSafe AI launched Jev on September 15, 2026, calling it the first "System One model", a category they coined. You send it program state plus a set of typed questions and it returns typed answers with probabilities, in one parallel pass, with nothing to parse. It comes from Diogo Almeida, who co-invented RLHF and InstructGPT at OpenAI, with $40M led by DCVC.
The mismatch it's aimed at is worth naming. LLMs produce text for humans to read. When you need a judgment your code will consume, you're coercing a text generator into structured output and then parsing it back into something you can depend on.
What is it, more precisely?
Three question types, mixable in a single call:
- Choice: pick one option from a list. Returns the choice, a probability distribution, and a confidence estimate.
- Score: rate the state on a rubric. Same return shape.
- Noul: is this statement true? Returns a probability between 0 and 1.
Each question is evaluated in parallel and in isolation against the same state, so adding questions barely moves the response time and there's no context rot between them. Training method is Reinforcement Learning for Calibrated Decisions (RLCD).
flowchart LR
accTitle: One Jev call, three parallel questions
accDescr: A single state is evaluated against three independent typed questions at once. Choice returns an option, Score returns a rating, Noul returns a probability. All three are evaluated in parallel against the same state.
state["state<br/><i>one block of context</i>"]
q1["<b>Choice</b><br/>one of N"]
q2["<b>Score</b><br/>rubric"]
q3["<b>Noul</b><br/>true / false"]
out["typed values<br/>+ probabilities"]
state --> q1 --> out
state --> q2 --> out
state --> q3 --> outThe numbers TypeSafe publishes: 40x to 200x faster than frontier LLMs, $0.042 per million input tokens with output free, responses in roughly 70 to 500ms, and about 68% accuracy on their own four-workflow benchmark, near mid-tier LLMs at a small fraction of the cost.
The design advice that comes with it matters as much as the model. Keep each question atomic: the kind of gut-check a knowledgeable person makes in seconds. If a judgment needs extended reasoning or weighs independent factors, decompose it and combine the results in your own code. Instead of "rate this pitch," ask separately about market size, feasibility and differentiation, then weight them with a formula you control. When priorities shift you change a coefficient rather than rewriting a prompt.
What you do with the answer is the architecture:
flowchart LR
accTitle: The confidence cascade
accDescr: Jev returns a decision with a confidence estimate. High confidence lets code act directly. Medium confidence triggers a confirmation or a request for more information. Low confidence escalates to a frontier model or a human.
jev["Jev decision<br/>+ confidence"]
high["code acts"]
mid["confirm or<br/>gather more"]
low["frontier model<br/>or human"]
jev -->|"high"| high
jev -->|"medium"| mid
jev -->|"low"| lowWhat it isn't
- Not a chat model. No text, no explanations, no code generation. Wrong tool for all three.
- Not a replacement for your LLM. The intended shape is a cascade: Jev classifies and routes cheaply, code handles what it can, a frontier model takes the hard minority.
- Not hallucination-proof in the sense the marketing implies. TypeSafe says it can't hallucinate because it only selects from allowed answers. That's a guarantee about format. A model can pick a permitted option and be flatly wrong.
- Not calibrated for your data out of the box. The confidence numbers come from the returned distribution; where you set the thresholds is your problem, and you have to measure how often confident answers are actually correct.
What I wish it could do
Tell me when it's drifting. The whole architecture assumes I've tuned thresholds against a labelled sample, but the distribution of what flows through a production classifier moves constantly. I'd want the model, or the platform around it, to flag that today's inputs no longer look like the ones I calibrated on.
I'd also like a cheap way to get the reasoning when I want it. Not on every call, since that would defeat the point, but on the escalated minority, where knowing why confidence collapsed is exactly the thing I need in order to fix the question.
What I'm still not sure of
Whether 68% is good. It sounds low until you ask 68% at what, against what baseline, at what cost per decision. A cheap fast wrong answer that gets caught by a threshold may well beat an expensive slow right one. But that calculus depends entirely on what a wrong decision costs downstream, which is a thing only I can price, and the benchmark is TypeSafe's own.
Bigger question: whether "System One" is a real category or a positioning. Structured output and classification are things existing models already do adequately. The claim here is that a purpose-built architecture does them two orders of magnitude more efficiently, and that the efficiency unlocks a class of decision you wouldn't currently pay a model to make. That second part is the interesting bet, and it isn't settled by a benchmark. It gets settled by whether people find the decisions.