Vercel AI Gateway Adds Liquid AI's d1-ultra Decision Model

Vercel AI Gateway has added support for d1-ultra from Liquid AI. This multimodal decision model is designed for classification, routing, and scoring across text and images, including use cases like visual defect inspection and verifying damage claims against customer photos.

Unlike generative models, d1-ultra does not produce text; instead, it returns structured answers with probabilities. Developers can ask Boolean questions to get the probability of a yes/no answer, use Choice to select among named categories, or use Score to rate an ordered scale. The model allows multiple questions about the same input in a single request, enabling developers to use the resulting probabilities to decide when to automate an action or escalate to human review.

Designed for harder real-world decisions with an emphasis on calibrated probabilities, d1-ultra supports a 32,768-token context window. It is billed solely on input tokens, including image inputs, with no charges for output tokens. The model is accessible via the AI SDK, OpenAI-compatible Decisions API, and TypeSafe-compatible API using the model name liquid/d1-ultra.