Jev introduces a new shape of LLM - System One, aka Decision Models

Blog · Watch/Read · (gathered 2026-09-23)

• Simon Willison reviews TypeSafe AI’s “Jev,” a new “System One” or “decision model” that accepts text input but outputs floating-point numbers (confidence scores, probabilities, ratings) for “Noul” (Yes/No), Choice, and Score questions, rather than text.
• Jev is exceptionally fast and cheap ($0.042/million input tokens, output free), allowing parallel evaluation of multiple questions against a single “state” object, making it ideal for classification tasks like spam detection, labeling, and search reranking.
• The model’s “black box” nature, providing only numeric outputs without justification, raises significant concerns about bias (e.g., in ranking job applicants) and necessitates rigorous, structured evaluations and experiments, which are made feasible by its low cost.