Jev

A decision model, not a text generator

Jev is the first model released by TypeSafe AI, a San Francisco company that emerged from stealth in mid-September 2026 with a seed round led by DCVC. Released in early access, it is not built to write text. TypeSafe calls it a System One Model, meaning one that makes decisions. The developer supplies unstructured input (a support ticket, say, or a document) along with typed questions, and the possible answers to each are fixed beforehand. Rather than composing a reply token by token, Jev gives back all the answers together, and each carries a calibrated probability and a confidence score. TypeSafe says this means the model cannot make a type error.

What it is meant for

TypeSafe aims it at decisions made inside everyday software, such as sorting and routing items, scoring them, pulling out data, checking facts and acting as a guardrail. According to TypeSafe, a request takes between 70 and 500 milliseconds from start to finish. Training used a method that TypeSafe has named Reinforcement Learning for Calibrated Decisions.

Vendor figures and independent tests

TypeSafe claims Jev is 193.6 times faster and 444.6 times cheaper than frontier models, but the comparison rests on TypeSafe's own workflow evaluations. Those workflows were written by TypeSafe's team, and the score measures how closely Jev agrees with the averaged answers of two other vendors' models. The company itself notes that this set-up may carry bias, and it says it cannot yet show that its pricing is not subsidized.

Every, a publication, ran independent hands-on tests on the same day. In the first, 37 documents were each put to Jev with 21 questions, and it finished them in under 0.7 seconds. In another, it caught six of seven deliberately planted defects, where a general-purpose frontier model at high effort caught all seven, but Jev took a median 0.35 seconds per passage against 8.83.

Access

Access is through a waitlist. No context limit, license or release of weights has been stated, and the product is described as being at an early stage.