Liquid AI released d1-ultra on 15 October, a model that looks at a photo and answers a question about it with a probability instead of a paragraph. The company’s own blog post calls it “the World’s Best Decision Model.” The head-to-head results behind that title are Liquid’s, and the post does not point to anyone else’s measurement of them.

A “decision model” is a narrower thing than a chatbot. It accepts text, images or both along with one or more questions, then outputs a likelihood for every candidate answer in one pass and writes no text. The questions come in three forms: yes or no with a probability from 0 to 1, a choice among many labels with a probability for each, and a score on a scale. Liquid’s examples: a customer’s photo that has to support a damage claim, and a board on the production line that has to match its spec. The code sample asks one question of a production-line camera image: “Does this circuit board have a defect?”

Liquid says “better-calibrated” is the point. Calibration means the confidence matches the record: if a model says 90 percent across a thousand boards, about nine in ten of those boards should really be faulty. In inspection that matters more than raw accuracy. A model that is wrong but unsure can pass the board to a person. A model that is wrong and certain sends the bad board down the line. Liquid says most decision models are overconfident and that it trained d1-ultra specifically to fix that. The text of the post gives no calibration figure, no error rate and no comparison number for it, only the assertion.

The comparison is against Liquid’s previous d1 and a model called Jev 1.14, across 112 evaluations the post does not list. Liquid says d1-ultra beats Jev 1.14 on every task, with sharper calibration and a higher hit rate, and it says the price is “the same.” The post does not say who makes Jev 1.14, and the text carries no per-task margins. On the Artificial Analysis Decision Index, it adds, d1-ultra sits at least 8 points clear of every other decision model. That is the nearest thing to an outside reference in the post, though it does not say who submitted the model or when.

The most useful comparison is missing. Circuit-board inspection already has an established machine-vision industry behind it, and the post sets d1-ultra against no such system. It gives no defect-catch rate, no false-alarm rate, no board types and no speed figure. The defect example is a demonstration, not a result.

Input tokens are priced at $0.042 per million, and that is the whole bill. Liquid charges for input only, because the model generates no output tokens. Images count at 1.5 tokens per 32 by 32 pixel patch, so a 1024 by 1024 image is 1,536 tokens. By our arithmetic, that is about six thousandths of a cent per image, or roughly 15,000 images per dollar. Every question counts as a separate prompt that carries its images, so asking five questions of one photo bills the photo five times.

d1-ultra is on the Liquid API now. Liquid says Vercel and OpenRouter will carry it “soon,” with no dates. Liquid says the base is a bigger open-weight model that it heavily customized, though it does not name it, and the post never says d1-ultra’s own weights are available.

A team with a few hundred labeled images of its own can test the claim in an afternoon at this price. The number to compute is whether the model’s 90 percent answers are right 90 percent of the time, and that is the one number Liquid has not published.

Based on Liquid AI’s blog post “d1-ultra: Introducing the World’s Best Decision Model” of 15 October 2026