A new model outputs calibrated decisions instead of text

TechCrunch reports on Jev, a model from a startup called TypeSafe AI that works differently from a normal language model. Instead of generating text token by token, it returns probabilities over a fixed set of possible answers, what the company calls calibrated decisions. Because the outputs are predefined, the model cannot invent a fact that is not on the list, so hallucination in the usual sense does not happen. The founder is Diogo Almeida, an OpenAI researcher who worked on ChatGPT and on reinforcement learning from human feedback, and Jev is trained with a variant he calls reinforcement learning from calibrated decisions on synthetic data.

The pitch is aimed at a specific kind of work. For classification, routing, monitoring other agents, and automation steps that need a reliable yes or no with a confidence score rather than prose, Jev is meant to be faster and cheaper. TypeSafe prices output tokens at zero and meters input by the billion rather than the million. Early users cited in the piece report real gains: Vercel says it ran 5 to 18 times faster with better accuracy than its previous OpenAI model, and Bryo AI says it was 10 to 20 times cheaper than Gemini while still returning useful confidence scores. Demand was high enough that the API briefly buckled.

The idea is a reminder that not every task needs a chatbot. A lot of production AI is really structured decision-making dressed up as text generation, and a model built for decisions directly can be both cheaper and easier to trust.

Why it matters

If you run classifiers or routing logic on a general chat model today, a decision-first model like Jev could cut cost and remove a class of hallucination bugs. It is worth testing on the parts of your system that only need a labeled answer with a confidence number, not free text.