TypeSafe AI launches Jev, a structured decision-making model

TypeSafe AI, a startup led by Diogo Almeida—a researcher involved in the development of ChatGPT at OpenAI—has announced the launch of its first System One Model, named Jev. Unlike traditional Large Language Models (LLMs) that generate sequential text, Jev is designed specifically for automation and machine interaction. It provides structured, typed probabilistic decisions based on predefined schemas rather than natural language strings.

Jev utilizes a new training method called Reinforcement Learning for Calibrated Decisions (RLCD) to optimize for epistemically honest probabilities. This architecture allows the model to perform parallel sampling, resulting in response times between 70ms and 500ms, which the company claims is significantly faster than conventional frontier models. Because the model is constrained by a schema, TypeSafe AI asserts that it cannot produce type errors or hallucinations in the traditional sense, though it provides confidence scores to indicate the reliability of its judgments.

Economically, Jev is positioned as a highly cost-effective alternative for developers using AI for routing, classification, and decision-making tasks. The company has set the input price at $0.042 per million tokens and offers output at no cost. While Jev cannot generate prose or code, it is intended to function as a high-speed, intelligent decision-making layer within larger software workflows.