The model that returns a number and charges nothing for it

Pick the model that owns your industry. That’s the advice for putting AI on a specialized job right now, and it’s good advice. Work in finance, you reach for the model trained on finance. Healthcare, the one trained on clinical data. Specializing an AI means going deeper into a field, and the field you’re in decides which model wins.

The numbers behind that advice are strong. Domain-tuned models beat general ones by up to 40 percent on domain tasks [verified]. BloombergGPT, Bloomberg’s finance model, runs about 30 percent more accurate on finance work than a general model does [verified]. Polaris, a clinical model from Hippocratic AI, reports 99.38 percent accuracy on medical use cases [verified]. The research firm Gartner expects more than half of enterprise AI deployments to use domain-specific models by 2027, up from 1 percent in 2024 [reported]. The question for a product team stopped being “which general model is best?” and became “which specialist owns my domain?”

Then a model shipped that specializes without going deeper into any domain. On September 21, a company called TypeSafe AI released Jev. Jev doesn’t write. You give it text and it hands back a number: a category, a yes or no, a rating, a confidence score. It’s built to reach a decision, not to produce prose. And its pricing is odd. You pay for what you put in, about four cents per million tokens of input, the tokens being the chunks of text these models bill by. You pay nothing for what comes out. Output is free [verified].

Simon Willison, an independent developer whose blog is one of the most-read running commentaries on what AI can actually do, called Jev a “Decision Model,” a new shape of model rather than a fine-tuned version of a general one [reported]. It isn’t a finance model or a healthcare model. It’s a decision engine that could serve any of them.

Specializing an AI doesn’t only mean going deeper into a domain. It can mean doing a different kind of thinking across all of them. For decision work, that’s a prior question the “which domain” framing never asks.

Watch where the two split. When the job is to write something domain-appropriate, draft the clinical note, summarize the trading position, the domain model wins and Jev has nothing to say. Fluent, correct domain output is the whole task, and that’s what domain training buys. But a lot of the work isn’t writing. It’s deciding. Is this transaction fraud? Where should this ticket route? How sure are we? Those come back as a yes, a category, a score. A decision rather than a paragraph. A model built to decide is a different thing from a model built to write in a finance vocabulary.

The price is the giveaway. A finance model can’t give its answers away, because the answer is the product. Its domain accuracy is the thing you’re paying for, and you can’t charge nothing for the thing you sell. Jev gives its output away free. That free output is the tell: the model isn’t competing on domain accuracy at all. Its value sits in the decision itself, and the words it would take to spell that decision out cost nothing because they were never the point.

The obvious objection is that these two aren’t rivals. They combine. The natural next product is a decision model tuned for a domain: Jev’s shape pointed at fraud rules, or clinical triage. If the two multiply instead of colliding, then Jev doesn’t break the “pick your vertical” advice. It adds a second dial. The domain still matters, and you’ve gained a knob rather than lost a map.

That’s true for the market and false for the advice, and the advice is what teams act on. The moment you need a second dial, “which vertical owns my domain?” stops being the question and becomes one question, scoped to writing. Because decisions are domain-specific too. Fraud classification is as domain-bound as work gets, specific rules, specific patterns, and it’s a pure decision. It sits fully inside the domain-model pitch and fully outside what a domain model is good at. The advice quietly assumed a specialized task means generating a specialized answer. Deciding is a different job from writing one.

Here’s the sharper miss. The people who argue hardest for domain models wrote down, in advance, the one thing that would change their minds: a single platform that matches vertical performance across domains without training a separate model for each one. Jev is exactly that shape. One architecture, every domain, no per-domain training. But it slips past the letter of the test. It doesn’t match a finance model at finance. It does a different job the finance model was never doing. The alarm was set for a challenger that beats the specialists at their own game. What arrived doesn’t play that game. It takes a whole class of work, decisions, off the specialist’s plate and prices it at zero. The test was written to catch a challenger coming straight down the domain axis. This one came in sideways.

Be precise about what’s proven. Jev exists, returns numbers, and charges nothing for output. That much is confirmed [verified]. That it’s a genuinely new category of model, and not a general one wearing a narrow output, is Willison’s reading, from a single blog post the day of the launch, with no second source yet [reported]. The broader idea that intelligent systems naturally split by kind of thinking rests on two research papers about how models train, not on any market [reported]. So this is one product, not a movement. It doesn’t overturn the case for domain models, which stands on verified production margins at named companies. It does something narrower and enough. It shows that case was mapping one axis of specialization and calling it the whole map.

So picture the actual choice. You’re sizing a high-volume classification job: fraud gates, ticket triage, confidence scoring across a domain. The domain camp tells you to pick the specialist that owns your field. The cost camp tells you to pick the cheapest model tier that clears your quality bar. Both are sorting models by a price you can put on them, accuracy or tokens. Jev answers a question neither one asked. Your task is a decision, so before you choose a domain or a tier, you’re choosing a kind of thinking. And the model for it charges you nothing for the answer.

Whether that’s a real third way to specialize or just a clever product has a clean test, and the test is the price. If decision models keep giving their output away, they aren’t competing on the axes the domain and cost frames measure, and “which specialist owns my domain?” was only ever the right question for the writing half of the work. If the next one charges for its answers the way a domain model does, it was a vertical all along and the old map holds. Watch what the next one charges for its output. The price will tell you before the benchmarks do.