TypeSafe AI released Jev, the first model in a new class the startup calls System One, aiming to replace slow, generative chat reasoning with fast, type-safe decisions that software can call directly instead of parsing free-form text.
The company was founded by Diogo Almeida, who says he helped build the instruction-following methods that became the research behind ChatGPT during his time at OpenAI. He argues that despite the enthusiasm around chat models, general-purpose language systems leave a gap for software that needs a decision rather than a conversation. System One models train with a technique TypeSafe AI calls Reinforcement Learning for Calibrated Decisions, optimizing for probabilistic, verifiable answers instead of the human-preference rewards behind typical chatbots. Jev takes unstructured program state as input and returns typed, probability-scored outputs rather than generated text, a tradeoff TypeSafe AI frames as removing hallucinated or malformed responses by construction.
TypeSafe AI frames the appeal as speed and cost. Jev answers System One-style queries in 70 to 500 milliseconds, against 3 to 329 seconds the company clocks for existing frontier language models on comparable tasks. According to TypeSafe AI, input tokens cost $0.042 per million and output is free during early access, and its published workflow evaluations weigh Jev's answers against the averaged output of GPT-6 Astra and Fable 5.1 on internal automation tasks. The company says the evaluation set was not built to flatter its own model, though it acknowledges its own team chose both the comparison models and the scoring method.
The tradeoffs are real. Jev gives up open-ended text generation entirely, and its parallel-sampling architecture only works for tasks with a fixed, defined set of possible answers rather than free-form conversation. TypeSafe AI is opening early access today and pulling developers off its waitlist, and it says it wants feedback on which decisions teams most need automated.
Whether a non-generative, type-safe model class catches on will depend on whether verifiable-decision workloads, like routing, scoring, and classification inside existing software, turn out to be common enough to justify a separate model family from the chat-style systems most AI budgets are still built around.
