TypeSafe released the Jev model on September 15, positioning it as a decision-making layer between code and LLMs. The system uses a typed API with three primitives: Noul for boolean decisions, Choice for selecting from a predefined list, and Score for probability distribution across a scale of 2–10.
Jev operates by receiving a state and a set of typed questions, returning answers with associated probabilities. The architecture supports speculative fan-out, allowing the system to process multiple branches simultaneously. As of September 18, the model version 1.13.0 was priced at $0.042 per million input tokens, with free output tokens.
The model is accessible via direct API, Vercel AI Gateway, OpenRouter, and Cloudflare Workers AI. Usage limits include 64k tokens per request and 1200 requests per minute. TypeSafe does not train Jev on user data, though specific retention policies for standard accounts remain undisclosed.
Technical analysis of the Jev-based browser agent shows it relies on predefined action spaces rather than generating code. In a test case involving flight searches, Jev accounted for approximately 52.59% of the total 7073 ms execution time. The system requires external verification to ensure task completion accuracy.
Comparative experiments suggest a trade-off between Jev’s zero-shot flexibility and the speed of specialized local classifiers. While Jev provides a structured decision interface, its performance depends on the quality of the provided state and the calibration of decision thresholds by the developer.
