Simon Willison reports that TypeSafe AI unveiled Jev last week, the first model in a category it calls 'System One models'—Willison and others prefer the term 'decision models.' Instead of generating text, Jev accepts a text or semi-structured 'state' object and returns floating point numbers: confidence scores for yes/no ('Noul') questions, probability distributions across multiple-choice options, or scores along a defined numeric range.
Pricing is unusual: Jev charges only for input tokens, at $0.042 per million, with output free—cheaper than OpenAI's GPT-5 Nano at $0.05 per million. Questions are evaluated in parallel against a single document, so sending many questions costs roughly the same as sending one.
Willison flags practical uses: spam detection, labeling, prioritization and ranking, and search reranking, where an inexpensive method like BM25 surfaces candidates and Jev scores them for relevance against the original query.
He also raises a concern specific to this format: Jev is a harder black box than a standard LLM, since it returns only a number with no justification. In one test, Jev rated Cupertino as the 'best' Bay Area city and East Palo Alto the worst on a 'Good city?' yes/no prompt, illustrating how bias can hide inside an unexplained score. Willison explicitly warns against using Jev to rank job applicants for this reason, and argues structured evals matter more here than for typical LLM projects—though he notes Jev's low cost makes running large batches of test prompts affordable.
The release has already spawned community experiments and open-weight recreations, including a project called Kev built on Qwen 3.5 in 0.8B, 4B and 9B sizes, plus a benchmark (JevBench) comparing decision models. Jev launched under a week before this report.