TypeSafe AI introduces System One Models and Jev for software decisions

TypeSafe AI has emerged from two years of stealth with the release of Jev, its first ‘System One Model’, a model class the company says is designed for software decisions rather than conversation. Founder Diogo Almeida announced the early-access release on 15 September 2026. Almeida previously worked at OpenAI on methods that helped language models follow instructions, according to his announcement.

TypeSafe says it built a separate stack around a new model architecture, a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. Jev accepts unstructured state but returns typed, probabilistic decisions that software can consume directly. The company contrasts that with language models that generate strings which applications must parse and validate. It also says every decision includes a confidence estimate.

The central promise is operational, not conversational. TypeSafe positions Jev for classification, routing, scoring, extraction, branching and other ‘smart if-statements’ inside automated workflows. Its published price is $0.042 per million input tokens, or $42 per billion, with output tokens listed as free. The company reports end-to-end response times of 70 to 500 milliseconds and says Jev is 40 to 200 times faster for comparable System One-shaped queries. Its homepage presents a narrower demonstration with claims of 193.6 times faster and 444.6 times cheaper.

Those figures are vendor-reported results, not independent benchmarks. TypeSafe says its workflow evaluations use the same code workflow for each model and compare outputs with averages from GPT-6 Astra and Fable 5.1. It also acknowledges that the workflows were made by people on its model-capabilities team and that the language models were run through a wrapper designed to produce compatible structured decisions. The company says its zero-type-error claim follows from schema matching, while its broader reliability claims still need real-world testing.

For developers, the release points to a different way of integrating AI: constrain the output space so surrounding software can act on uncertainty instead of interpreting prose. If the early-access results hold outside TypeSafe’s own evaluations, small, fast decision models could make AI practical in high-volume, latency-sensitive workflows. The open questions are whether Jev generalises across domains, how its confidence calibration performs on unfamiliar cases, and how the service handles privacy, uptime and failure recovery in production.