Ex-OpenAI Researcher's TypeSafe AI Ships 'Jev,' a Model That Only Picks From a Fixed List

A former OpenAI researcher has built a model that refuses to write anything at all. Diogo Almeida's new company, TypeSafe AI, is releasing a system called Jev that answers only by selecting from a set of options supplied by the developer — no prose, no chat, no free-form output. The company says responses start at roughly 70 milliseconds, with token pricing kept very low.

That is a deliberate inversion of how most vendors now sell language models. The dominant pitch is breadth: one system that drafts emails, writes code and holds a conversation. Jev instead narrows the job to classification, the everyday task of routing a support ticket, labeling a record or flagging a transaction. Almeida's bet is that plenty of production workloads never wanted generated text in the first place — they wanted a label, and they were paying for fluency they did not need.

The speed and cost figures are the sales argument. At 70 milliseconds, a classification can sit inside a request path where a chat model's multi-second latency would force engineers to bolt on queues, caches or a cheaper fallback. Low per-token pricing matters just as much for teams running millions of judgments a day, where inference cost, not accuracy, often decides what ships. TypeSafe is pitching Jev at that volume tier rather than at interactive assistants.

Jev's constraint also has a sharp edge, and the company is not hiding it. The model guarantees only that its output will be one of the permitted choices; it does not guarantee the choice is correct. A wrong classification is still a wrong classification, just a neatly formatted one. The design removes a class of failure familiar to anyone who has parsed model output in production — malformed JSON, stray commentary, an invented category — but it leaves accuracy entirely in the hands of the underlying prediction.

That distinction matters because it separates two problems that get bundled together. Schema compliance is an engineering problem, and TypeSafe is claiming to solve it structurally, since a model that can only emit valid options cannot emit an invalid one. Correctness is a modeling problem, and it remains as hard as ever. Buyers evaluating Jev will need to measure it the way they measure any classifier: precision, recall and error rates on their own labeled data.

The launch lands as the industry argues over whether general-purpose models will absorb every specialized use case or whether focused systems will keep winning on latency and price. TypeSafe is making the second case with a model that does less on purpose. What to watch next is whether developers treat fixed-option classification as a feature worth adopting on its own, or simply as a setting they expect their existing model provider to offer. Almeida's answer should arrive in the form of adoption numbers, not benchmarks.