Liquid AI's d1-ultra decision model arrives on Vercel AI Gateway

Liquid AI's d1-ultra, a multimodal model built for decision-making rather than text generation, is now available through Vercel's AI Gateway. According to the Vercel blog, developers can call it under the model name liquid/d1-ultra and put it to work on classification, routing and scoring tasks involving text and images.

A model that answers with probabilities

d1-ultra does not produce free-form prose. It returns structured answers with probabilities, an approach intended for workloads where the output has to feed directly into application logic.

Vercel outlines three question types. A Boolean query returns the probability of a yes or no answer; a Choice query picks among named categories; and a Score query rates an input along an ordered scale. Several questions about the same input can be bundled into a single request.

The emphasis on calibrated probabilities is the distinguishing feature. Vercel positions the model as suited to difficult real-world decisions, and the returned confidence levels can act as a threshold: high-confidence results trigger automated actions while borderline cases go to a human reviewer.

Suggested use cases include inspecting goods for visual defects and checking damage claims against photos submitted by customers, both situations where an image must be judged against criteria and turned into a dependable verdict.

Context window and billing

The model supports a 32,768-token context window. Its billing is unusual: charges apply to input tokens only, including image input, with no fees for output tokens. Since d1-ultra emits probabilities and categorical picks rather than generated text, output volume is minimal by design, and the pricing reflects that.

Integration with AI Gateway

On the platform side, d1-ultra can be reached through the AI SDK, an OpenAI-compatible Decisions API and a TypeSafe-compatible API, and Vercel's examples pose the same question across all three. Decision calls appear in Gateway logs and count toward budgets alongside requests to other models, so teams can monitor usage and cost in one place.

Vercel also notes that AI Gateway passes provider pricing through without adding a markup and takes no platform fee on inference, including on Bring Your Own Key requests. Developers who want to experiment before writing code can try the model in Vercel's model playground.

Why it matters

Most model availability announcements add one more text generator to the pile. d1-ultra belongs to a different category: a model that outputs probabilities for classification and routing, which is what many production workloads such as content triage, moderation, fraud checks and image inspection actually consume.

For developers already on AI Gateway, the addition means decision tasks can run through the same unified API, logging and budgeting as their generative models, and input-only billing makes costs easier to estimate in advance. Calibrated confidence matters in practice as well: when the reported probability reliably reflects real-world accuracy, teams can set automation thresholds deliberately and keep humans in the loop for genuinely ambiguous cases.

What the announcement does not cover is how d1-ultra's calibration compares with alternatives in the field. Developers can form their own view through the playground and the decision models documentation Vercel points to.