Jev introduces a new shape of LLM - System One, aka Decision Models

TypeSafe AI has put out Jev, the first example of what it calls a “System One model.” Simon Willison says “decision model” is the better label, and that seems to fit: Jev takes text in and sends back floating point numbers instead of paragraphs of text.

The company is pitching it as a kind of frontier-intelligence function call. You feed it a state object — a string, an array of strings, or named values — and then ask it questions about that input. The answers come back as numbers for yes/no checks, choices, or scored ranges, along with confidence data. Multiple questions can be packed into one request, and they’re evaluated in parallel, so the promise is speed as well as simplicity.

Pricing is part of the pitch too. Jev charges only for input tokens, with output free, and its first model is priced at $0.042 per million tokens. That puts it below OpenAI’s GPT-5 Nano, which Willison cites at $0.05 per million tokens. For the kind of experimentation this model invites, that matters: running lots of prompts is supposed to cost just a few cents.

Willison’s main use-case argument is pretty practical. Jev looks well suited to classification work like spam detection, labeling, ranking, and prioritization. He’s also been trying it for search reranking, where a cheaper system finds likely matches and Jev scores them for relevance against the original query.

But the thing that makes Jev neat is also the thing that makes it uneasy. It pushes even further into black-box territory. If a model says something is spam or rates a job applicant, the result is just a number. No explanation, no obvious trail. That makes bias harder to spot, and Willison says evals and structured experiments become even more important. The community has already started playing with it anyway, from fake chat systems to left-pad and 2048, and open-weight recreations are already appearing.