Liquid AI releases d1-ultra, a decision model that returns probabilities

Liquid AI's d1-ultra answers yes/no, label and scale questions about text and images with probabilities, billed at $0.042 per million input tokens.

Liquid AI released d1-ultra on October 15, 2026. It is a multimodal decision model that reads text, images or both and returns a probability for each answer. It does this in one forward pass and generates no output tokens. If your pipeline asks a chat model for JSON just to get a yes/no or a label, this is a cheaper, typed alternative worth testing. Liquid's examples include checking a damage claim against a photo and spotting a circuit-board defect.

The API keeps d1's three question types: noul (yes/no), choice (one label from many) and score (a position on a scale). Liquid says d1-ultra runs on a larger open-weight base than d1. It also says it was trained specifically for calibration, because "most decision models are overconfident." Liquid reports that the model beat d1 and Jev 1.14 across 112 evaluations. Crypto Briefing reports that it scored 65 on the Artificial Analysis Decision Index, against 57 for Jev 1.14. That index only began public operations on October 13, and the outlet notes that Liquid's internal evaluations haven't been independently verified.

Pricing is $0.042 per million input tokens, with images at 1.5 tokens per 32×32 patch. Each question is billed as its own prompt, including all images, so five questions about one photo cost five times the image tokens. Superpower Daily flags this as the cost detail to model.

Liquid's post says Vercel and OpenRouter access is coming "soon". Both already have listing pages. Vercel's page shows a 32,768-token context window and a $0.04 input rate, and OpenRouter's page shows $0.042. Neither page confirms whether image input works on those gateways. If you want something cheaper, base d1 lists at $0.02 on OpenRouter.

Before you trust the calibration claims, test them on your own labeled data.

Sources