TypeSafe AI has released Jev, the first model in what it calls System One Models, a class built to return typed, probabilistic decisions instead of generated text. Founder Diogo Almeida describes a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Jev is available in early access.
The central claim is that many automation steps are narrow decisions that do not need free-form strings at all. Because the possible outputs are defined in advance and all options are scored in a single parallel query, the model cannot emit type errors and every answer carries a probability that software can branch on. TypeSafe lists input pricing of $0.042 per million tokens and end-to-end response times of 70 to 500 milliseconds.
The timing matters because TypeSafe positions Jev for routing, scoring, extraction, and guardrail steps that currently run through slower general-purpose LLM calls. The company cites 193.6x faster and 444.6x cheaper results. Those results do not establish accuracy against ground-truth labels, since the workflow evals score agreement with the average of GPT-6 Astra and Fable 5.1 and were built by TypeSafe's own team.
For engineers, the practical test is whether a decision step in an existing workflow can be expressed as a fixed set of typed options with a confidence threshold. Independent comparisons against small fine-tuned classifiers and constrained decoding, plus calibration checks under data shift, would show whether the claimed gains hold outside TypeSafe's workflows.