According to Simon Willison, writing on 21 September 2026, TypeSafe AI last week unveiled Jev, described as the first example of a category the company calls "System One models" (Willison, citing Maggie Appleton, prefers the label "decision models"). Willison reports that Jev accepts text inputs but, instead of returning text, returns floating point numbers corresponding to categories, yes/no questions, ratings and associated confidence scores. Users compose a "state" object—a string, array of strings, or name-value pairs describing a record such as an article or customer—and send it with one or more questions, which are evaluated in parallel. On pricing, Willison writes that Jev charges only for input, with output free, at $0.042 per million input tokens, which he notes is cheaper than OpenAI's GPT-5 Nano at $0.05 per million. He describes suitable uses as classification tasks such as spam detection, labeling, prioritization and search reranking. Willison raises concerns that Jev is a "black box": it returns only a number, offering no explanation of which signals drove a decision, and he flags bias risks, stating he hopes nobody uses it to rank job applicants. He recommends structured evals given the low cost. He also notes community projects replicating the approach on open-weight models, including "Kev" built on Qwen 3.5, and a "JevBench" benchmark.
- Jev returns floating point numbers with confidence scores instead of text output
- Priced at $0.042 per million input tokens, with output free
- Suited to classification tasks: spam detection, labeling, ranking, search reranking
- Willison flags black-box opacity and bias risk, urging structured evals
What it means for you
This is a model built to answer yes/no, rate, or categorize—it hands back a number with a confidence score rather than a paragraph. That makes it fast and cheap for jobs like sorting spam, tagging records, or ranking search results. The catch: it tells you what it decided but never why, so you can't audit its reasoning, and any hidden bias stays hidden.
Try this
If you already run any classification task (routing support tickets, flagging spam, tagging content), it's worth a small test—the cost is a few cents to run hundreds of examples—but only after you build a labeled test set to check whether its answers are actually right.
Who should care
Developers and technical teams doing high-volume classification, tagging, or ranking who want something cheaper than a full text-generating LLM.
Skip this if
You don't build software, you need decisions you can explain (loans, hiring, anything with legal or fairness stakes), or you're not doing classification-style work at scale. Absolutely do not use this to rank job applicants.
Sources: Simon Willison — read the original