Overview
- TypeSafe AI removed Jev’s waitlist and gave new users a $5 starter credit roughly equal to 120 million tokens while publishing an input-only price of $0.042 per million input tokens with outputs free.
- Jev is designed to accept a program-provided state plus a bounded question and return structured answers—Choice, Noul (probability), or Score—that code can act on as a reusable decision module.
- The public launch prompted heavy adoption and integrations, caused an early API outage, and produced rapid open-source experiments and demos including a Jev plus GPT-6 Astra Minecraft Ender Dragon run reported to cost under $1.
- Community tests and TypeSafe’s own notes report reliability and calibration issues such as sensitivity to option order, inconsistent probabilities on repeated runs, poor handling of arithmetic, counting and dates, and vulnerability to adversarial inputs.
- The model’s low cost and speed could change how teams build agent loops by offloading many small judgments to a dedicated module, but adopters should guard against miscalibrated probabilities and run-time instability before using Jev for high-risk automation.