Jev: The AI Decision Model Cutting Costs 100x and Reshaping the Agent Stack

TypeSafe AI launched Jev on September 15, a specialized AI model built not for text generation but for decision-making — the routing, scoring, and branching logic that consumes the bulk of AI agent compute. Priced at $0.042 per million input tokens with zero cost on output, it returns structured choices, scores, or yes/no probabilities in under half a second. Within 72 hours, Vercel, Cloudflare, LangChain, and Langfuse had all integrated it into their platforms — a remarkably fast institutional endorsement.

DCVC led a $40 million seed round backing founder Diogo Almeida, a former OpenAI researcher who helped develop the instruction-following research underlying ChatGPT. He spent two years building in stealth alongside co-founders Erik Gafni and Sasha Sheng. The company named the model after 19th-century economist William Stanley Jevons, whose paradox holds that efficiency gains in resource use tend to increase total consumption rather than reduce it — a dynamic TypeSafe is deliberately betting on for AI.

Technically, Jev accepts a block of application state plus typed questions and returns a Choice (up to 255 options), a Score (ordered scale), or a Noul (calibrated yes/no probability). All questions in a request are evaluated in parallel, and schema compliance is guaranteed — the model cannot hallucinate an option that wasn't offered. TypeSafe trained it using a method called Reinforcement Learning for Calibrated Decisions, designed so that confidence scores reliably track accuracy, enabling automated action above a threshold and human escalation below it.

Real-world benchmarks are striking. Writer Mike Taylor processed 777 content-quality judgments across 37 documents in under 0.7 seconds for roughly a quarter of a cent. In a head-to-head against Claude Fable 5.1, Jev ran 25 times faster at approximately 1/580th the cost, catching six of seven planted defects versus Fable's perfect seven. One developer used Jev to clear 9,081 product-matching pairs — a queue his team had abandoned in June because frontier model pricing made the task economically irrational — for 32 cents in 13 minutes. That example captures the core thesis: at $15 per million output tokens, certain tasks simply don't get built. At zero, they do.