Stop paying for output tokens with Jev and these "decision models

Imagine an LLM that actually knows how to shut up and just give you a number. TypeSafe AI just dropped Jev, which they call a "System One" model, though "decision model" fits better. Instead of hallucinating a three-paragraph essay when you just wanted a yes or no, Jev takes text and spits out floating point numbers. It's basically a frontier-intelligence function call: you feed it unstructured state, and it gives you typed probabilistic decisions.

The pricing is the only thing that makes sense in the current market. Jev only charges for input tokens—output is free. At $0.042 per million tokens, it's actually cheaper than GPT-5 Nano, which sits at $0.05 per million.

How do you actually use this thing?

You don't "chat" with Jev. You send a "state" object (a string, an array of strings, or name-value pairs) and attach one or more questions. Since it evaluates everything in parallel, you can cram as many questions as the context window allows without the latency spiking.

You're limited to three types of queries:

• Noul questions: (Short for Bernoulli, because why make it easy?) You make a statement, and it returns a float between 0 and 1 representing its confidence that the statement is true.
• Choice questions: You provide a set of options, and it returns a probability distribution across those options plus a confidence score.
• Score questions: You define a numeric range with descriptions, and it places the input somewhere on that scale via a floating point score.

Where does this actually work?

If your task is basically classification, this is the move. Spam detection, labeling, or ranking—things where a prose response is just noise. I've been playing with it for search reranking: use a cheap algorithm like BM25 to grab 100 potential matches, then let Jev score those 100 candidates for relevance against the query. It's significantly faster than asking a standard LLM to "rate these 100 items" and hoping it doesn't lose count.

The "Black Box" problem

Here is the annoying part: we are moving backward into total opacity. With a standard LLM, you can at least beg it to "explain its reasoning" (even if it's just lying to you). Jev gives you nothing but a number. If it flags a document as spam or ranks a lead as "low quality," you have zero visibility into which specific signals triggered that decision.

This makes bias a massive concern. If you're using a floating point number to rank job applicants, you're basically trusting a black box to handle human careers without a single word of justification. It's efficient, sure, but it's a blind leap of faith.