TypeSafe AI claims System One task intelligence comparable to existing large language models while operating roughly two orders of magnitude faster and cheaper. Its parallel sampler produces every output in a single query instead of generating tokens sequentially, and the company reports an input price 238 times lower than Claude Fable 5.1.

Response latency of 70 to 500 milliseconds contrasts with frontier language model responses that range from 3 to 329 seconds. Jev returns type-safe structured outputs with guaranteed schema matching, and because it does not generate strings it avoids the text hallucination and output parsing problems that come with prose generation. Its calibration is designed so higher confidence tracks higher accuracy, and on workflow evaluations the model occupies the Pareto frontier across nearly two orders of magnitude.

TypeSafe AI frames each order-of-magnitude drop in the cost of intelligence as unlocking orders of magnitude more use cases. The company targets AI powered workflow decisions, map-reduce operations over large data sets, real-time applications, and AI verification and guardrails, and argues that the token market could eventually become as broad as the energy market.

The Register cautions that token-demand assumptions remain unsettled and places near-term adoption in sorting customer service problems, while flagging open questions about the round. TypeSafe expects pricing to fall over time rather than be subsidized, and describes itself as early in the life of Jev with more products still in the pipeline.

TypeSafe AI is a San Francisco artificial intelligence lab developing machine-native, composable models intended for use by software rather than by people. The company introduced its first public product, Jev, a System One Model, in early access on September 15, 2026.

Jev accepts unstructured state, either a JSON object or a plain string, and returns typed probabilistic decisions rather than generated text. Its answers rely on question primitives such as Choice, Score, and Noul, and every decision carries a calibrated confidence estimate so surrounding code can set thresholds for autonomous action or escalation.