Liquid AI extended d1 to image-based decisions on October 5, following an earlier experimental text release. The model takes a situation and a set of questions, then returns probabilities for answers such as yes or no, a category, or a position on a scale.
That format can handle tasks such as routing support messages or deciding whether an image shows a defective part. It does not generate a free-form explanation. Liquid’s examples also use d1 to choose browser actions and decide which tool outputs to retain in a coding agent’s context.
The company reports lower costs than larger chat models on six application tests, but each application was run only once per model. The comparison used list prices without prompt-cache discounts, so the ratios should not be treated as a general cost ranking.
Vision is available through Liquid’s API and playground. The announcement lists Vercel and OpenRouter access as text-only, with vision to follow. Billing covers input tokens, including images, and each question is charged as its own prompt with the shared input included again.