TypeSafe AI released Jev, a System One model for decisions. It returns typed choices with probabilities and generates no text. Jev processes the input once and scores all options in parallel. It supports choice, score and yes/no question types. TypeSafe reports 70 to 500 ms latency per call. It charges $0.042 per million input tokens with output free. Jev is calibrated, with an expected calibration error of 0.0588 on one benchmark. It fits tasks where answers are selected from a list. It struggles with arithmetic and step-by-step reasoning. It is suited to frequent, hot-path decisions. It works with text-only state inputs.
What changed
Autoregressive models generate text answers token by token in a loop.
What it unlocks
Selecting from a fixed list of options with calibrated confidence and no text generation.
- 70 to 500 ms per call
- $0.042 per million input tokens
- 0.0588 expected calibration error
- 255 options max per choice
What you need to act on it
- a list of options provided in advance
- tasks with structured output
Sources