Overview
- TypeSafe launched Jev after two years in stealth with $40 million in seed backing, positioning the model as a “System One” component that returns typed decisions and calibrated probabilities instead of natural‑language text.
- The company removed the waitlist and opened Jev to public sign‑ups on September 20, offering new users $5 in starter credits and a pricing model that charges only for input tokens at about $0.042 per million while making output tokens free.
- TypeSafe claims Jev is optimized for embedded decision tasks with typical latencies around 70–500 milliseconds, a roughly 32,000–64,000 token read‑only context window, and a new training method called Reinforcement Learning for Calibrated Decisions (RLCD).
- Independent testers have found adversarial and out‑of‑scope failure modes that can steer Jev’s outputs or produce inconsistent results, and TypeSafe’s documentation advises empirical calibration on domain data plus conservative human‑in‑the‑loop thresholds.
- Developers have quickly integrated Jev across major platforms such as Vercel, Cloudflare, Netlify AI Gateway, LiteLLM, and OpenRouter, and the model’s low per‑call cost could drive broad use for tasks like triage and gating while raising new safety and validation needs.