TypeSafe AI Launches Jev, a 'Decision Model' That Outputs Only Numbers, Priced at $0.042/M Input Tokens

TypeSafe AI has launched Jev, the first model in a category the company calls "System One models"—critics and community members, including researcher Maggie Appleton, prefer the term "decision models." Rather than generating text, Jev accepts text or semi-structured data as input and returns floating-point numbers representing classifications, confidence scores, or ratings.

How Jev works

Users send a "state" object—a string, array of strings, or set of name-value pairs describing a document, customer, or record—along with one or more questions. Jev supports three question types:

• Noul questions (named for the Bernoulli distribution): yes/no statements answered with a confidence score between 0 and 1
• Choice questions: the model selects from provided options, returning a probability distribution across all choices
• Score questions: a numeric range with described levels, returning a floating-point score along that scale

Questions submitted against a single state are evaluated in parallel, so sending many questions takes roughly as long as sending one, according to TypeSafe AI.

Pricing

Jev charges $0.042 per million input tokens with output free entirely—a departure from standard LLM pricing, which typically charges more for output than input. That input rate undercuts OpenAI's GPT-5 Nano, currently priced at $0.05 per million input tokens, making Jev one of the cheapest inference options available as of this writing. TypeSafe AI has not disclosed a context window size for Jev.

Use cases and early experiments

Developer Simon Willison, who tested the model extensively, describes decision-model framing as useful for classification tasks: spam detection, labeling, prioritization, and ranking. Willison also tested Jev for search reranking—using BM25 to fetch 100 candidate matches, then having Jev score each for relevance to a query.

The developer community has produced unconventional experiments in the days since launch, including jevchat (a token-by-token chat model built entirely from yes/no questions), jev-leftpad (a reimplementation of the left-pad npm package using choice queries), and jev-2048 (an implementation of the 2048 sliding-puzzle game). A JevBench benchmark has also emerged to compare Jev against open-weight alternatives, including Kev, a project from developer Jared Palmer that fine-tunes Qwen 3.5 into 0.8B, 4B, and 9B parameter decision models.

Black-box concerns

Jev's output format raises transparency questions. Unlike conventional LLMs, which can be prompted to explain their reasoning—however unreliably—Jev returns only a number, with no visibility into which input signals drove the decision. Willison flagged bias risk explicitly, cautioning against using Jev for tasks like ranking job applicants, where a single opaque confidence score could encode unexamined bias. In one informal test, Jev rated Cupertino, California highest and East Palo Alto lowest on a "Good city?" yes/no question—an outcome with no accompanying explanation.

What this means

Jev's release signals a bet that not every AI task needs natural-language output—many production use cases (moderation, ranking, tagging) only need a number, and pricing that number at a fraction of a cent could make LLM-based classification viable at scales where token-generation costs were previously prohibitive. But the opacity trade-off is real: without interpretability tooling, teams adopting Jev-style decision models will need rigorous evals to catch bias before it reaches production, and TypeSafe AI has offered no benchmark scores of its own to validate accuracy claims.