A former OpenAI researcher has launched a new class of AI model aimed squarely at automation rather than conversation. Diogo Almeida, who worked on the instruction-following methods that became the research behind ChatGPT, spent two years in stealth building TypeSafe AI. On 15 September the company unveiled its first System One Model and a public early-access model called Jev. The pitch is blunt: chat models have been superhuman at talking for years, so where is the automation?
Jev is built for fast, structured decisions that software can call directly. The company says it reaches frontier-level intelligence on System One tasks while running 40x to 200x faster, with end-to-end responses of 70ms to 500ms instead of seconds. It gives up string generation entirely, producing only pre-defined, type-safe values with calibrated confidence scores attached. TypeSafe calls it a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.
The economics are the headline for builders. Input tokens cost USD 0.042 per million, with output tokens effectively free, and the company claims Jev cannot hallucinate because it has no free-form text to invent. Training uses a method TypeSafe calls Reinforcement Learning for Calibrated Decisions, optimising for honest probabilities rather than human preference. The approach trades the flexibility of chatbots for reliability inside pipelines, and it will be judged on whether that bet pays off.