TypeSafe launched Jev, a new AI model that fundamentally differs from standard large language models by returning predefined "typed" decisions rather than free-form text responses. The service functions as a structured classifier that accepts inputs and provides defined outputs using a "System One" decision model, and has rapidly attracted developer interest since its recent release. This constraint-based approach aligns with the principle that limitations often drive creative problem-solving.

Jev operates through three query types—Choice, Score, and Noul (a truthfulness probability score)—and requires developers to define the schema and candidate options beforehand rather than using open-ended prompts. Performance is notably fast, with response times between 150 and 620 milliseconds, and significantly cheaper than frontier LLMs at $0.0011 per decision and $0.042 per million input tokens with no output token charges. Early adopters have deployed applications ranging from spreadsheet urgency classification and virtual clothing fitting to gaming optimisations and CPU emulation, with a community site called Jevable collecting prototypes that showcase both practical productivity tools and experimental projects.