Jev was released September 15. The model accepts text or semi-structured state and returns typed probabilities: a yes/no confidence, a distribution over choices or a score. It does not write prose. At $0.042 per million input tokens with free output, it is designed for cheap, repeated decisions inside agent systems. TypeSafe documentation, Simon Willison’s technical account
The developer response supplies useful secondary signals. Developer experiments appeared almost immediately—Jev playing 2048, reproducing left-pad, reranking search results and being coerced into character-by-character chat. TypeSafe briefly lost API availability under demand, and integrations arrived through Vercel, Cloudflare, LangChain and Langfuse. Yesterday Willison released an llm-typesafe plug-in that makes the three query types inspectable from his CLI. The pattern is a primary launch followed by several independent artifacts and a credible observer explaining why developers care.
The appeal is also the risk. Jev supplies a number without a rationale. It is weak on numbers, dates and adversarial content, according to its own documentation, and a confident score can conceal bias more efficiently than a paragraph can. Cheap evaluation makes large test suites affordable; it does not make them optional.