Liquid AI Unveils Decision-Making Model

Boston-based startup Liquid AI, spun out of MIT, released d1-ultra on October 15, 2026 — the largest multimodal "decision-making model" that returns answer probabilities for typed questions in a single pass without generating tokens.

Compiled by Sergey KostenchukPublished 2026-10-21Updated 2026-10-21

2026-10-21 Research

🤖 Liquid AI Unveils Decision-Making Model

On October 15, Liquid AI (Boston, spun out of MIT) released d1-ultra — the largest multimodal "decision-making model." Input is a situation (text, image, or both) and typed questions (yes/no, label, position on a scale); output in a single pass is the probabilities of the options without token generation. According to the company, d1-ultra outperforms participants in the Artificial Analysis Decision Index by an average of 8 points; the probabilities are calibrated: 80% corresponds to approximately 80% accuracy.

🌍 This is for fast structured decisions: defect classification, ticket routing, insurance, moderation. Calibrated probabilities provide an adjustable threshold, and you only pay for input tokens ($0.042 per 1 million).

👤 Test for free — in the d1.liquid.ai sandbox; available via Liquid AI API, OpenRouter, and Vercel AI Gateway.

Source 1: https://www.liquid.ai/blog/d1-ultra Source 2: https://openrouter.ai/liquid/d1-ultra

Sources