TypeSafe AI opened early access to Jev, a model that returns decisions instead of text

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, opened early access to a model called Jev on September 15, 2026. The company calls it a System One Model, a class built for the decisions software makes inside itself rather than for conversation with a person. Jev never produces a string. It returns values whose type and range were fixed before the call, and it attaches a confidence score to each one.

What the model returns

A language model answers in text, so the program around it has to parse that text and validate it before acting. Jev removes that step. The set of possible answers and their structure are declared in the query itself, the model fills them in, and every answer arrives with a calibrated probability beside it.

Because the shape is fixed in advance, the model cannot return a field that does not exist or a value outside the declared set. TypeSafe presents that as the practical meaning of its claim that Jev cannot hallucinate, and the claim is narrower than it sounds: the answer is always well formed, which is not the same as always correct. Sampling runs in parallel rather than token by token, and that is where the company locates its speed advantage.

Price and response time

Input costs $0.042 per million tokens, which the company also writes as $42 per billion. Output tokens are not billed. For comparison, the same post puts the input price of current language models between $0.20 and $10 per million tokens, with output typically about five times the input price.

TypeSafe reports end-to-end response times of 70 to 500 milliseconds, measured on its own machines on the West Coast of the United States. It says the same calls take frontier language models 3 to 329 seconds.

Where the numbers come from

The multiples on the home page, 193.6 times faster and 444.6 times cheaper, are drawn from four evaluation workflows the company published alongside them. Every model receives the same workflow, and the reference answer is the average of what GPT-6 Astra and Fable 5.1 return.

The company states the caveats itself, which is uncommon enough in a vendor post to be worth naming. It says its own model capabilities team built the workflows. It says averaging two competing models biases the reference toward those two. It says the published multiples sit at the high end of what it expects in production. The hallucination rates it plots for language models come from OpenRouter traffic, while its own zero is derived from the schema guarantee rather than measured.

What it is not for

Jev answers questions whose possible answers are known before the call. It classifies, routes, scores, extracts, and branches, and the number of choices in a single decision is capped at 255. Above that the call splits in two: the options are scored first, then one is chosen. Anything that has to be written, whether a reply, a document, or code, still needs a language model.

The offer, then, is a replacement for the small classification calls buried inside an application, not for the assistant a user talks to. Whether the published multiples hold on someone else's workload is not answerable yet, because no independent measurement has been published.