Liquid AI has launched d1-ultra, which returns decision probabilities instead of generated prose. It accepts text, images or both and supports yes/no questions, label choices and scores. Examples include circuit-board inspection and photo-backed damage claims.
The model is available through Liquid’s API at $0.042 per million input tokens, with no output tokens. Each question is billed as its own prompt containing its text and every image, so additional questions repeat the billed input. Images are tokenized. Vercel and OpenRouter availability is described as forthcoming, not established by this release.
Liquid claims improved accuracy and calibration against Jev 1.14 in its evaluations. This does not guarantee business-decision reliability or establish open weights for d1-ultra; it is distinct from October 7’s open d1 models.
Bounded decision calls are becoming a separate agent layer for routing and repeated checks.