Liquid AI has released d1-ultra, a multimodal decision model that scores 65 on the Artificial Analysis Decision Index. The new model reads text, images, or both, and answers typed questions with a probability for each possible answer in a single forward pass without generating output tokens. Available now on the Liquid API at $0.042 per million input tokens, it is designed for developers handling classification, routing, and scoring tasks.
Built on a larger open-weight backbone than its predecessor, d1-ultra evaluates multiple typed questions against the same input context simultaneously. The system supports three question types including yes/no queries, label choices, and score rubrics, making it ideal for moderation, triage, and LLM-as-judge replacements. According to Liquid AI, the model leads across all eight categories in its evaluation suite and ranks first on the independent Artificial Analysis Decision Index, outperforming competing systems like TypeSafe's Jev.
Developers can access d1-ultra through the Liquid API with support from the Vercel AI Gateway and OpenRouter arriving soon. Billing covers input tokens only, with image costs calculated based on pixel patches, while parameter counts and downloadable model weights remain undisclosed. This release matters because it offers developers a cost-effective, token-free API solution for high-throughput decision-making tasks like content moderation and system routing.