Liquid AI's d1-ultra answers decision tasks with probabilities, not generated text

Liquid AI released d1-ultra this week, a multimodal model for a niche that frontier labs rarely touch: decision tasks. Instead of generating text, the model reads text, images or both in a single forward pass and returns a probability for each possible answer. The company calls it a state-of-the-art decision model, and it is priced at $0.042 per million input tokens on the Liquid API.

Decision models: a different contract

A decision model answers questions about a situation with a probability for each possible answer. d1-ultra takes unstructured input, text, images or both, plus one or more questions, and returns the probabilities without generating any tokens. It supports three question types, the same set as the earlier d1:

• noul: a yes/no question, answered with a probability from 0 to 1.
• Choice: one label from many, with a probability per label.
• Score: a position on a scale.

The practical angle: if your workload is classification, validation or scoring, you are currently paying for output tokens you throw away. The decision API bills on input only.

The claims, with numbers

According to the company, d1-ultra is based on a larger open-weight model with a series of Liquid-specific customizations, and the combination with Liquid AI's training pipeline brings significant gains, especially on knowledge-related tasks. Across 112 evaluations, Liquid AI says it outperforms its previous d1 model and Jev 1.14 on all tasks at the same price, and it leads every other decision model by at least 8 points on the Artificial Analysis Decision Index.

One detail stands out: calibration. Most decision models are overconfident in their responses, Liquid AI says, and d1-ultra was specifically trained to produce more reliable probabilities. If you route decisions off these numbers, calibration matters as much as raw accuracy.

"The world's best decision model" is the company's claim. The index gives the state-of-the-art line a verifiable basis, but it is one benchmark, on a narrow comparison set of decision models only. Some will call it hype, and that is a fair read.

The numbers that matter for a budget

Billing is input tokens only, with no output tokens, because the model generates none. Images count as input tokens at the same rate as text: 1.5 tokens per 32 by 32 pixel patch, so a 1024 by 1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all its images. The rate is $0.042 per million input tokens.

d1-ultra is live on the Liquid AI API, and the company says it will be available through Vercel and OpenRouter soon.

If you have a classification-shaped workload, defect detection, claim checking, moderation, it is worth modeling it as decision questions instead of asking a chat model to classify and parsing the answer afterwards. The cost shape is simpler, input only, and the output is already numeric. The practical test is whether the calibrated probabilities hold on your own data. The d1 playground has demos running on d1-ultra, and it is the right place to start before wiring up the API.