Jev
The first public System One model from typesafe-ai, announced after two years in stealth and covered in jev-is-incredible. Classification, not generation: it takes application state plus a JSON shape and returns typed decisions with confidence values on every answer. A frontier intelligence function call rather than a chat partner, meant to be integrated like a library.
Specs and claims as reported in the video:
- 32k context window. No vision yet, which the presenter treats as the feature whose arrival unlocks frame triage and stronger computer use.
- 70 to 500 ms classification. About 4 cents per million input tokens, output tokens free.
- Zero structured-output and tool-call format violations in TypeSafe’s charts, the property the company calls type safety. See structured-outputs.
- Statelessness between calls, visible in the Doom demo as left-right thrashing, since every frame is a fresh decision with no memory of the last one.
- Named for Jevons’ paradox: making power cheaper raised coal consumption, and making judgment cheaper is expected to raise how often it gets used.
Theo’s line for what it is not: a judge between reasoning-model outputs, or a compactor of conversations. It is also no replacement for the LLMs in a coding harness. See system-one-models for the category.