TypeSafe AI released Jev, a model that returns only typed structured values and cannot hallucinate

TypeSafe AI came out of stealth on 15 September and released Jev, the first model in a new class the company calls a System One Model. The company was founded by Diogo Almeida, who worked on the instruction-following methods behind ChatGPT at OpenAI. Jev is in early access. The launch post is on the TypeSafe blog, and the technical writeup Latent Space ran the same day.

What the model does

Jev does not generate text. It takes unstructured input plus a schema that names the fields and the possible values, and returns a typed record with a calibrated probability on each field. TypeSafe describes this as "unstructured state in, typed probabilistic decisions out". The intended jobs are the ones people already wire language models into: classify a message, route a request, score a candidate, extract fields from a document, branch on a fuzzy rule.

Two design choices follow from that scope. Jev samples every field of the output in parallel in one query, rather than one token at a time. And because the possible outputs are defined by the schema in advance, TypeSafe says a type error is mathematically impossible and the model cannot invent a value that is not in the enum. There is no free-text field for a hallucination to hide in.

The numbers TypeSafe published

TypeSafe says Jev matches existing frontier LLMs on the tasks it was built for, and comes in 40 to 200 times faster and about two orders of magnitude cheaper. Its pricing is $0.042 per million input tokens, with output tokens free. It puts end-to-end response time at 70 to 500 milliseconds against 3 to 329 seconds for the frontier models it compared against.

On its own four-workflow evaluation set, TypeSafe reports Jev is up to 193.6 times faster and 444.6 times cheaper than the reference. The reference is the average of GPT-6 Astra and Fable 5.1 outputs, run through a TypeSafe wrapper that constrains those LLMs to structured decisions. TypeSafe states these caveats itself: the workflows were made by its own model-capabilities team, the reference biases toward OpenAI and Anthropic, and speedups are on the higher end of what to expect in production.

The trade the model is asking a team to make

Jev does not do anything a language model does with strings: no chat, no code, no free-text answers, no reasoning trace to read. If the job is "make a fast, structured decision inside a running program", that is fine and is the entire pitch. If the job needs a written explanation of the answer, it is the wrong tool.

Set aside the speed number for a moment: the load-bearing claim is the schema guarantee. Every existing LLM has to be wrapped in a validator that catches malformed JSON, unknown enum values, or a hallucinated field, and every wrapper is one more piece of code that runs on every call. A model that cannot emit an invalid record removes that wrapper. The piece to verify against a team's own data before trusting the model on customer-facing decisions is whether the calibrated probabilities are honest, because that number decides when the code branches without a human in the loop.

Source

• Introducing System One Models and Jev, TypeSafe AI, 15 September 2026.
• Jev: a "System One Model" that only decides, classifies, routes, and scores, Latent Space, 16 September 2026.