TypeSafe AI unveils System One and Jev: decisions instead of text generation

TypeSafe AI has introduced System One, a new class of models that don't write text but instead immediately produce a finished structured decision with probability estimates. Company founder Diogo Almeida previously worked at OpenAI on the methods that formed the basis of ChatGPT; the team spent two years developing its project behind closed doors.

The first release is the Jev model, opened up in early access. Instead of generating token by token, Jev returns all answers in a single request, and the possible options and their structure are defined in advance. That means the format makes the usual formatting errors impossible in principle, and hallucinations in the familiar sense are ruled out — although the model can still pick the wrong option. Every answer comes with calibrated probabilities and a confidence estimate.

By TypeSafe's own measurements, Jev reaches a level comparable to modern LLMs on its own tasks, but is two orders of magnitude faster and cheaper: $0.042 per million input tokens, with no charge for output at all. Response latency is 70–500 ms, against 3–329 seconds for frontier models. The company's homepage cites figures of 193.6x for speed and 444.6x for cost.

The approach works differently: instead of training on human preferences or verifiable rewards (RLHF and RLVR), TypeSafe uses RLCD — reinforcement learning for calibrated decisions — plus a parallel sampler that computes all outputs in a single request. The names are deliberate too: the model class references the "fast" thinking from Kahneman's book, and the name Jev references William Jevons.

The main application is the role of a "fuzzy if" inside ordinary code: classifying, routing, scoring, extracting data, and branching logic where rigid rules are too brittle. Another scenario is checking and constraining other models: evaluating, judging, catching jailbreaks. Fast responses make it possible to embed such calls in latency-sensitive applications.

The company spells out the limitations itself. The figures come from internal workflow evaluations, benchmarked against the average of responses from GPT-6 Astra and Fable 5.1 — meaning the comparison is skewed toward OpenAI and Anthropic solutions. The authors acknowledge that the benchmarks may have been chosen in their own favor and that in reality the gains will be smaller. There is no independent verification: access to Jev was opened in early mode, and developers are being taken off the waitlist gradually.

Image:

Translation editor: Салтанов А.