TypeSafe AI launches Jev, a model that answers in numbers

TypeSafe AI has released Jev, the first model in a class the company calls "System One models". Jev takes text as input but does not answer in text; it returns floating point numbers for categories, yes/no questions and ratings, together with confidence scores. Only input is billed, output is free. The first model costs $0.042 per million input tokens, against $0.05 for OpenAI's GPT-5 Nano.

The API takes a "state" object made of a string, a list of strings or name-value pairs, plus as many questions as fit into the context window. Jev answers yes/no questions with a value between 0 and 1, choice questions with a probability distribution across all options and score questions with a value on a given scale. Questions run in parallel, so many take about as long as one. According to the Jev 1.13 documentation, the model is weak on numbers, dates and "adversarial content"; Simon Willison has added it to his LLM command-line tool through the llm-typesafe plugin.

Why it matters: Anyone filtering spam or routing messages to a team in automation workflows gets a number from Jev instead of free text that a downstream step first has to interpret. Because Jev does not explain its decisions, it stays unclear which signals tip the balance. That makes testing all the more important, and according to Willison hundreds or thousands of requests cost only a few cents. He advises against using it to rank job applicants, because the number can hide bias that nobody sees.