Ex-OpenAI researcher's Jev outputs decisions, not text, and wins over developers

TypeSafe AI, a startup founded by Diogo Almeida — an early OpenAI researcher who worked on ChatGPT and helped invent reinforcement learning from human feedback — has released Jev, a transformer-based model that deliberately does not speak. Instead of generating text, Jev returns probabilities, which the company brands "calibrated decisions," for tasks whose possible answers the developer defines up front. According to TechCrunch, developer interest was strong enough that the company briefly lost the ability to serve API traffic after launch.

A model that cannot hallucinate

The design choice has practical consequences. Because the output space is fixed in advance by the user, TechCrunch notes, the model cannot hallucinate the way language models do: there is no free-form prose in which to invent facts. It is also cheap to run — output tokens cost nothing, and input tokens are metered by the billion rather than by the million.

TypeSafe calls Jev a "System One model," aimed at fast intuition rather than step-by-step reasoning. The name nods to William Stanley Jevons, the 19th-century economist whose paradox holds that as a commodity gets cheaper, total consumption of it tends to rise. Almeida's stated bet is that the falling cost of decisions will spread intelligence across many small pieces of software rather than concentrating it in a few large chatbots.

What 227 milliseconds looks like

The most vivid illustration of the speed gap is a Pong demo built by Ably Labs that reached the front page of Hacker News. Four lanes run the same game with the same serve and the same question posed to a different model in each lane; each answer advances the ball one step, so a model's latency literally sets its pace.

On the demo's recorded run, Jev averaged 227 milliseconds per decision, 4.40 decisions per second, with a p95 of 400ms and 47 decisions inside the first 12 seconds. The chat models managed two or three decisions in the same window: per the demo, Google's Gemini 3.8 Flash averaged 3,167ms, Anthropic's Claude Haiku 4.5 averaged 2,483ms, and OpenAI's GPT-5.6 Sol averaged 3,529ms. The demo's authors chose Pong because it is the Atari game chat models handle worst, and because a game loop cannot pause to wait for a slow reply.

Early adopters and their numbers

TechCrunch quotes developers already swapping Jev in for chat models. Pranit Sharma, a software engineer at Vercel, said his team had used OpenAI's ChatGPT Luna 5.6 as a classifier reviewing agent commands for safety; replacing it with Jev produced results five to 18 times faster, with greater accuracy. Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for classifying business emails and found Gemini slightly more accurate but 10 to 20 times more expensive — with Jev's real probability scores being the more useful property for automating workflows.

Armin Ronacher, CTO of Earendil, which builds the open-source model harness Pi, frames the model as shifting judgment to the developer: if a decision comes back near 50 percent, treat it as a coin toss and disregard it; at 95 percent, act on it. He also points to model routing — cheaply predicting which model a given workload needs — and Almeida suggests using Jev to watch LLM agent traces and intercept jailbreaks, a job that gets expensive when done with another LLM.

An unverified architecture

How Jev works under the hood is not public. Almeida is tight-lipped about the architecture, TechCrunch reports, and outside observers suspect it is built on top of an open-weight LLM. What he does disclose is that it is trained exclusively on synthetic data, using a technique he calls "reinforcement learning from calibrated decisions," and that the internal lab producing that data now makes up half the company. More versions, in new modalities, are planned, and Ronacher expects competitors to appear now that the category's usefulness is visible. Almeida, for his part, rejects the frontier-lab label, saying he would rather TypeSafe's product be intelligence than hype.

Why it matters

Nearly every serious AI deployment today routes through chat models: general and capable, but priced and paced for conversation. Jev is an early and prominent bet that a large class of automation work — classification, safety checks, routing, monitoring — is better served by a model that emits a number in milliseconds rather than an essay in seconds. If that bet holds, intelligence becomes cheap enough to scatter through ordinary software, which is exactly the Jevons-style outcome Almeida is aiming for. The caveats are real: the architecture is unverified, the comparisons come from vendor-adjacent demos and individual developer anecdotes, and the model is days old. But the traction suggests non-language decision models are now a category others will have to take seriously.