TypeSafe AI launches System One Models and Jev

TypeSafe AI, a startup that spent two years in stealth, has announced its first System One Model: a class of frontier models built to make fast, structured decisions that software can use directly. The first public model, Jev, is available in early access. The announcement was written by founder Diogo Almeida, who previously worked at OpenAI on the methods that made language models useful at following instructions.

From strings to typed decisions

Jev does not generate text. Its outputs are type-safe structured values whose shape and possible answers are fixed in advance, so the model cannot make type errors; the company says that guarantee is mathematical rather than empirical. Every answer arrives with calibrated probabilities and confidence scores. Almeida describes the model as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.

TypeSafe argues that the bottleneck for automation is not intelligence but the interface. A string can be a chat reply, code, a refusal or a hallucination, and any software that consumes it must parse and validate the result. Because the output is constrained, Jev is meant to slot into ordinary code as a fuzzy decision rule: classify, route, score, extract or branch where hand-written logic is too brittle.

RLCD, parallel sampling and published numbers

The stack rests on a new architecture, a parallel sampler and a training method the company calls Reinforcement Learning for Calibrated Decisions, in place of the reinforcement learning from human feedback and verifiable rewards used for chat models. The claimed difference is scale: Jev returns all outputs in a single query instead of generating one token at a time. TypeSafe publishes input pricing of $0.042 per million tokens, output tokens free, and end-to-end response times of 70 to 500 milliseconds, against 3 to 329 seconds for frontier models — a 40x to 200x speedup.

On its homepage the company cites figures of 193.6x faster and 444.6x cheaper, drawn from workflow evaluations in which every model receives the same compute graph and predictions are compared with the average of two reference models. TypeSafe notes that its own team built the workflows, that the reference answers lean toward OpenAI and Anthropic models, and that its results should therefore be treated with caution.

Demos and the name

The team demonstrated Jev playing Doom from structured game state at about ten queries per second and roughly $7 per hour, and a game of Wikiracing in which each step means choosing among hundreds or thousands of links; above 255 options the model uses a two-stage scheme that scores first and then makes an explicit choice. The model class is named after Daniel Kahneman's Thinking, Fast and Slow, while Jev is named after William Stanley Jevons, the economist whose paradox held that cheaper coal did not reduce coal consumption. TypeSafe is making the same bet on intelligence: every order of magnitude cut in cost unlocks new uses.