Diogo Almeida, founder of Type Safe AI, announced Jev, a new System One Model that achieves similar intelligence to existing LLMs on structured tasks while being 40-400x cheaper and 20-200x faster. After two years in stealth, Type Safe AI built a new stack with a novel architecture, parallel sampler, and Reinforcement Learning for Calibrated Decisions (RLCD).

Jev is optimized for structured outputs and cannot hallucinate. It takes unstructured state and returns typed probabilistic decisions. Unlike LLMs that generate strings sequentially and require parsing, Jev generates all outputs in a single query with calibrated probabilities.

Cost: input tokens at $0.042 per MTok ($42 per billion), output tokens free. Speed: end-to-end response time 70ms-500ms versus 3-329 seconds for frontier models. Confidence is always communicated, making Jev suitable for automation where human oversight is impractical.

Jev is available today in early access.