TypeSafe Jev: decisions with probabilities

TypeSafe AI came out of two years in stealth with Jev, a model that takes state and questions and answers with code-ready choices, scores and yes or no, each with a probability. Input costs 4.2 cents ($0.042) per million tokens, output is free.

In short

Jev does not generate free text. Answers are typed and each comes with a calibrated probability.
The answer arrives in 70 to 500 milliseconds; all outputs are computed at once, in parallel.
The speed and cost figures come from the company's own tests, and it lists the caveats itself.

Checked on1 October 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

A large part of the code around a language model is cleaning up after it. You parse the answer, check whether it is in the right format, run it again when it is not.

Jev offers to remove that part. It does not ask the model to write an answer. It asks it to choose.

The facts: on 15 September 2026 TypeSafe AI announced the first model in a class it calls System One, after two years in stealth. The founder is Diogo Almeida, who says he worked at OpenAI on the methods that led to ChatGPT. Jev takes unstructured state and a set of questions and returns typed answers - a choice, a score or yes and no - with calibrated probabilities and confidence; all outputs are generated at once, in parallel. It is trained with a method the company calls Reinforcement Learning for Calibrated Decisions (RLCD). The price is 4.2 cents ($0.042) per million input tokens, and output is not charged; the answer arrives in 70 to 500 milliseconds. Per the company, in its own workflow evaluations Jev is up to 193.6 times faster and 444.6 times cheaper than language models, with the average of GPT-6 Astra and Fable 5.1 serving as the reference. The model is in early access.

The caveats, written by themselves

The good part is that they listed them themselves. The tests are by their own team. The reference is the average of two outside models; in their words this tilts the reference answers towards OpenAI and Anthropic and, if anything, underrates Jev itself. Speed was measured mostly from laptops on the US West Coast, where their server is too. And the price may be subsidised, and they say so plainly.

A model that answers only with a choice from a list cannot invent an answer outside the list.

That is the real offer, more important than the speed. In their words Jev cannot hallucinate, and the logic is clear: not because it is smarter, but because it has nothing to hallucinate with. It can pick the wrong option, but not one outside the list. The price of the trick is just as clear: Jev does not write letters, does not summarise and does not chat.

If your code has a spot where a language model only decides yes or no - classifies, routes, scores a risk - try it there. Anywhere else it does not belong.

The visual is generated code art. No third-party images.

TypeSafe released Jev: a model that returns decisions with probabilities instead of text

The caveats, written by themselves

More on this

UK AI Security Institute: a Claude Mythos 5 agent tried to slip malicious code into a real project using fake identities

Open models passed half the traffic at Vercel: 56 per cent of tokens, but 14 per cent of the money

Cloudflare gives agents a workspace, a small model for decisions and a search that starts costing money in November