TypeSafe AI introduces Jev, a new shape of LLM known as System One, aka Decision Models

Summary

TypeSafe AI has launched 'Jev,' a decision model. The model accepts text input and returns typed probabilistic outputs (yes/no scores, category selections, numerical ratings) instead of generated text. The cost per input is just $0.042 per million tokens; output is free.

Why it matters

Jev repositions large language models (LLMs) as classification and scoring functions rather than text generators. This approach can make high-volume tasks like spam filtering, search re-ranking, and prioritization significantly cheaper and faster. Parallel question evaluation and zero output cost could fundamentally change how teams build structured decision pipelines on top of base models.

Highlights

• Output is free; input cost is $0.042 per million tokens. This price is lower than OpenAI's GPT-5 Nano model.
• Three question types (Noul/yes-no, Selection, Score) are evaluated in parallel. This allows many questions to be answered with the same latency as a single question.
• Within a week of Jev's launch, open-weight recreations (Kev based on Qwen 3.5) and the JevBench comparison emerged.

Critical take

Willison highlights that the return to opaque, black-box scoring lacking transparency is unsettling. The model offers no explainability or reasoning; there is only a decimal number. This makes bias auditing and trust-building much harder compared to a text-generating LLM that you can guide through its reasoning. The city ranking example (Cupertino at the top, East Palo Alto at the bottom) is a small but striking reminder: 'cheap and fast' does not mean 'fair.'