Jev’s System One Models Offer Fast, Cheap Alternative to LLMs

TypeSafe AI recently unveiled Jev, a novel model category they call “System One models” (or “decision models,” as many observers have noted). Unlike traditional large language models that generate text outputs, Jev returns structured data—specifically, floating-point numbers representing confidence scores for various categories, yes/no questions, and ratings.

How Jev Works

Jev accepts text or semi-structured data inputs (like articles, customer records, etc.) and allows users to pose multiple questions in parallel. The API supports three question types:

• Noul (yes/no) questions return a confidence score between 0 and 1
• Choice questions provide a probability distribution across options
• Score questions map inputs to predefined numeric ranges with descriptions

This structured output makes Jev particularly well-suited for classification tasks like spam detection, content labeling, and prioritization—applications where clear decisions matter more than nuanced explanations.

Performance & Cost Advantages

Jev offers significant advantages over traditional LLMs:

• It’s exceptionally fast due to its focused output format
• Input costs are just $0.042 per million tokens (cheaper than OpenAI’s GPT-5 Nano)
• Output is free, as Jev only charges for input data

Practical Applications

Early adopters have found Jev useful for:

• Search reranking: Scoring candidate results from initial searches
• Content categorization: Automatically assigning labels and tags
• Prioritization engines: Ranking tasks or items based on defined criteria
• Decision support systems: Providing structured input for automated workflows