TypeSafe's Jev returns typed decisions, not text

TypeSafe AI announced a model on September 15, 2026 that cannot write a sentence. Jev takes the messy state of a running program as input and returns a structured, typed value. The company's announcement says it produces that whole value in one pass, instead of one token at a time. TypeSafe calls this a System One model, a class it says is built for automation rather than chat.

The trade is explicit. Jev gives up text generation entirely, and in exchange the company claims latency of 70 to 500 milliseconds and an input price of $0.042 per million tokens with output free.

What a System One model does differently

An ordinary language model writes left to right, choosing one token at a time, and a structured answer is something you coax out of that stream. TypeSafe inverts it. Jev is given a structure up front and fills it in, all at once.

The outputs carry calibrated probabilities, meaning each decision comes with a confidence number meant to match how often it is actually right. TypeSafe says it trained the model with a method it calls Reinforcement Learning for Calibrated Decisions, rather than the RLHF approach used for chat models. The stated goal is confidence numbers that track real accuracy.

Artificial Intelligence News reports one concrete mechanism behind the scale of those choices. For selections with many options, up to 255 of them, Jev scores candidates in two independent stages before picking.

The founder is Diogo Almeida, who worked on ChatGPT research at OpenAI. Both the company and Gigazine say TypeSafe spent two years in stealth before this launch.

The numbers, and who produced them

Every figure below comes from TypeSafe or from reporting of TypeSafe's claims. No independent benchmark exists yet.

Claim	Figure
Latency	70 to 500 milliseconds end to end
Speed versus frontier models	40x to 200x, per TypeSafe; Gigazine reports 20x to 200x
On multi-step workflows	193.6x faster and 444.6x cheaper
Input price	$0.042 per million tokens, against $0.20 to $10 for chat models
Type errors	0%
Demo	A DOOM bot running about 10 queries per second, at $7 an hour

That DOOM figure is the most useful one, because it is a rate rather than a ratio. Ten decisions a second at seven dollars an hour is a number you can compare against your own workload.

The headline claim needs care. TypeSafe says hallucination is mathematically impossible, and the mechanism it describes is type safety: the output must conform to the structure you defined. That guarantees the shape of an answer, not its truth. A decision can be perfectly typed, validate cleanly, and still be wrong, which is what the confidence score is there to surface.

What this means for developers

The honest way to evaluate this is as a replacement for a specific call, not for your model. Picture a step in a pipeline that asks an LLM to choose one of a fixed set of options and return JSON. That is the shape Jev is built for. Classification, routing, ranking and tool selection fit. Anything that has to produce prose does not, because the model cannot.

Price the comparison on your own traffic before believing the multipliers. Free output tokens change the arithmetic most for workloads where the response is small and the input is large, which is exactly what a decision call looks like. A summarization workload would see none of that benefit.

Treat the vendor benchmarks the way this site treated Salesforce's CRM Bench numbers: as a claim to reproduce, not a measurement to plan around. The 193.6x figure is specific enough to be testable, so test it on one real decision in your stack.

Two practical limits before you schedule anything. Access is early, through a waitlist, with documentation at docs.typesafe.ai, so this is not something you can drop into production this week. And the calibration promise is the piece to verify yourself: log the confidence numbers alongside outcomes for a week, then check whether a 0.7 really means seventy percent.