TypeSafe AI unveils Jev, a decision-focused model that returns numeric outputs instead of text

TypeSafe AI last week unveiled Jev, its first example of a new category of model it calls "System One" models, which Simon Willison and others refer to as "decision models." Unlike conventional LLMs that return text, Jev accepts text or semi-structured data as input and returns floating-point numbers corresponding to categories, yes/no answers, ratings, and associated confidence scores.

Jev is priced only for input tokens at $0.042 per million, with output provided at no additional cost. This pricing model undercuts OpenAI’s GPT-5 Nano, which is priced at $0.05 per million input tokens. The model’s input format, called a "state," can be a string, array of strings, or set of name-value pairs describing a record such as an article or customer.

Users can pose three types of questions to Jev: "Noul" (Bernoulli) questions for yes/no statements, returning a confidence score between 0 and 1; choice questions that return a confidence score and probability distribution across provided options; and score questions that return a floating-point score along a user-defined numeric scale.

Jev evaluates multiple questions in parallel against a single input, enabling batch classification tasks such as spam detection, label suggestion, prioritization, and ranking. Willison notes its utility for search reranking, where BM25 retrieves candidates and Jev scores their relevance to the original query.

The model’s design emphasizes speed and cost efficiency, with Willison observing that running hundreds or thousands of experimental prompts costs just a few cents. However, he cautions that Jev’s black-box nature—returning only numeric outputs without explanatory text—complicates interpretability and bias assessment, warning against high-stakes applications like applicant ranking without rigorous evaluation.

Within days of Jev’s release, the community produced open-weight recreations such as Kev, which uses Qwen 3.5 to create 0.8B, 4B, and 9B variants, and community benchmarks like JevBench to compare "Jev-class decision models." Creative experiments have also emerged, including a chat-like interface, an implementation of the left-pad function, and a version that plays the 2048 sliding puzzle game.