The Information: Jev maker TypeSafe AI is in talks to raise over $1B at a $10B+ valuation

The Information reports that startup TypeSafe AI is in talks to raise more than $1 billion in a new round at a valuation above $10 billion. Nine days earlier, on September 15, the company launched Jev — its first System One model — together with a $40 million seed round; PitchBook data puts the valuation of that round at $200 million.

TypeSafe AI is based in San Francisco and was founded in 2024. It is led by chief executive Diogo Almeida, a former OpenAI researcher credited as a co-creator of InstructGPT. His co-founders are Erik Gafni and Sasha Sheng, and the seed round was led by DCVC.

What Jev does

Jev is not a language model: it neither writes nor chats. It takes unstructured state — text, JSON or an array of strings — with typed questions, and returns a choice from a predefined set, a score on an ordered scale, or a yes/no probability between 0 and 1. Every answer carries a calibrated confidence value, so software can decide whether to act or hand the case to a person.

Because the answers are fixed before the call goes out, the model cannot return a value outside the schema; TypeSafe presents that as the elimination of hallucination and type errors. The company quotes response times of 70 to 500 milliseconds and a price of $42 per billion input tokens, with output not metered. Its own evaluations report Jev as 193.6 times faster and 444.6 times cheaper than frontier LLMs, while noting that real-world gains are likely to be smaller.

Early traction and open questions

Vercel says Jev became the fastest-adopted model in the history of its AI Gateway: within 24 hours of launch it was used by nearly 13% of paid teams — twice the share of the GPT-5.6 family. The model is free through the gateway until September 25, so its October share will say more than the first day did.

Jev remains in early access: TypeSafe has not named production customers and does not disclose revenue. The company says the model is transformer-based and trained only on synthetic data; outside observers suspect it is built on an open-weight LLM. The reported talks are a bet on a narrower idea — that routing, scoring and approval inside agent pipelines deserve a specialized model rather than a general one.