TypeSafe says its new model can't hallucinate, because it never writes a sentence
Declare the answers you'll accept and Jev fills them in with probabilities, which is why TypeSafe says it can't hallucinate and charges nothing for output tokens.
On September 15, 2026, TypeSafe AI released its first model. Jev doesn't write text at all. You declare the questions and the shape of each answer in advance, and it hands back typed values with a probability on every one.
Diogo Almeida founded the company. He says he helped build the methods that made language models good at following instructions at OpenAI, and that the work ended up as the research behind ChatGPT. Then two years in stealth.
- Released
- Sep 15, 2026
- From
- TypeSafe AI, founder Diogo Almeida
- Input tokens
- $0.042 / MTok
- Output tokens
- FREE (too cheap to meter)
- Response time
- 70ms-500ms
What surprised me is the housekeeping. Almost every claim in the post carries a block headed Nuance underneath it, and one of those blocks says the number above it isn't empirical.
What it gives up
A language model writes one token at a time, and your code parses whatever comes back. Jev evaluates every declared question at once and returns the answers already typed, so there's no string to parse. (The class name comes from Daniel Kahneman's fast and slow thinking. Jev itself is named for the economist William Stanley Jevons.)
Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.
Because the possible outputs are fixed before the call, TypeSafe says the model can't hallucinate. There are limits on what you can ask for. One choice can hold up to 255 options, and above that the model scores them and then picks in a second pass.
The model never makes type errors
The numbers, and what TypeSafe says about them
End to end, a Jev call takes 70 to 500 milliseconds. The post puts frontier models at 3 to 329 seconds for the same kind of query. Input runs $0.042 per million tokens, or $42 per billion, and output tokens are free.
The headline figures on the company's home page are 193.6x faster and 444.6x cheaper, and the post says where they come from: a set of workflow evals where every model gets the same workflow and is graded against the average of GPT-6 Astra and Fable 5.1. TypeSafe says that reference biases the results toward OpenAI's and Anthropic's models and probably understates DeepSeek's. It also says its own capabilities team built the workflows, so some bias could exist.
On the hallucination chart, the LLM figures come from OpenRouter and the Jev figure is a zero the company put in by argument rather than measurement. Granted, the argument is that a schema match is guaranteed.
Where you can call it
Vercel added Jev to its AI Gateway, reached through an experimental evaluate API in AI SDK 7. The uses Vercel lists are the small decisions inside an agent loop: which tool to call next, and whether to keep going, retry, ask the user or stop.
Jev evaluates all declared questions in parallel and returns typed answers plus probabilities directly.
Jev is early access only. TypeSafe says it's bringing developers off the waitlist as quickly as it can.