TypeSafe AI emerged from stealth on Sept. 15, 2026, with $40 million in seed funding led by DCVC and early access to Jev, its first System One model.
Jev is not a chat model. Developers feed it program state and structured question types such as Choice, Score, or Noul; it returns typed probabilistic values and confidence scores that software can consume directly. TypeSafe calls the training approach Reinforcement Learning for Calibrated Decisions, or RLCD, and says Jev cannot emit free-form text the way ChatGPT-style models do.
Co-founders are Diogo Almeida, a former OpenAI researcher and co-inventor of RLHF methods tied to ChatGPT, with Erik Gafni and Sasha Sheng. TypeSafe said the company was founded in 2024 in San Francisco. In company materials, Almeida said most intelligence should eventually live inside software rather than only serve people as the consumer. Early access is waitlisted at typesafe.ai.
Company claims on performance should stay labeled. The Business Wire release said Jev runs at less than 100 milliseconds of latency and is up to 100 times faster and less expensive than other frontier models. The company blog lists $0.042 per million input tokens, free output tokens, and 70ms-500ms end-to-end response times, and separately cites homepage workflow figures as high as about 194 times faster and about 445 times cheaper as higher-end gains on its own evals.
Independent coverage: The Register reported the typed-decision interface, the Doom demo, and the DCVC seed, while attributing the speed and cost multipliers to TypeSafe. The Register also said a "hallucination-free" pitch is not a fair comparison to natural-language models, because structured wrong answers remain possible. Dealroom corroborated the $40 million DCVC-led seed and called the round large for seed stage on its dataset. Windows Forum framed the performance numbers as vendor measurements, not third-party benchmarks.
What remains open is whether outside labs will reproduce the latency, cost, and reliability claims on third-party workloads, and how far typed System One decisions travel beyond early-access demos. Company speed and cost figures and third-party reporting should be weighed separately until independent benchmarks land.