Last week Type Safe AI unveiled Jev, their first example of a new category of model they call 'System One models'—also known as 'decision models'. Jev is an LLM variant that accepts text inputs but returns floating point numbers for categories, yes/no questions, ratings, and confidence scores. Type Safe describes Jev as a 'frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.' It's fast and cheap, charging only $0.042 per million input tokens—cheaper than Open AI's GPT-5 Nano at $0.05/million.

Jev lets users ask questions about text or semi-structured data by composing a state object and sending it to their API with multiple questions evaluated in parallel. The model supports three question types: Noul (yes/no confidence scores 0-1), choice questions with probability distributions, and score questions with numeric ranges. This approach is excellent for classification tasks like spam detection, labeling, and search reranking.

However, Jev represents a regression toward black box systems—it provides no explanations for decisions, raising concerns about hidden bias, particularly in sensitive applications like hiring. Extensive evaluation is crucial given the model's opacity.