Liquid AI releases d1-ultra decision model
Liquid AI has launched d1-ultra, a multimodal model that returns calibrated probabilities instead of free text, aimed at harder real-world checks such as circuit-board defect detection and insurance-claim verification.
Liquid AI announced d1-ultra, the successor to its d1 decision model. It is available now on the Liquid API at $0.042 per million input tokens. The company says Vercel and OpenRouter support is coming soon.
Decision models answer structured questions rather than generating open-ended text. d1-ultra supports the same three question types as d1: a yes/no question answered with a probability from 0 to 1, a choice of one label from many with a probability for each label, and a score on a scale. It accepts text, images, or both together.
Liquid says d1-ultra is built on a larger open-weight base model than d1, with "Liquid-specific heavy customizations." It does not name the base model. The company says most decision models are overconfident, so it trained d1-ultra specifically to produce better-calibrated probabilities.
According to Liquid, it tested d1-ultra against d1 and a model called Jev 1.14 on 112 evaluations. It says d1-ultra beat Jev 1.14 on every task, with higher accuracy and better calibration at the same price. Liquid also says d1-ultra leads all other decision models on the Artificial Analysis Decision Index by at least 8 points. The blog post as provided includes no individual benchmark scores.
Billing counts input tokens only. Images cost the same as text, at 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all its images. That means asking several questions about one image multiplies the cost.