Liquid AI releases d1-ultra decision model: calibrated probabilities with zero generated tokens, #1 on Artificial Analysis Decision Index (65), $0.042 per M input tokens

Background

Classification, routing, moderation and guardrail tasks are usually done by asking an LLM to generate text and parsing it, which is slow, costly and poorly calibrated. Liquid AI open-sourced small d1 decision models earlier in October and on Oct 15 shipped the hosted flagship d1-ultra for harder cases such as checking a circuit board against its spec or a damage claim against customer photos.

How it works

d1-ultra reads text, images or both and returns probabilities in a single forward pass with no generated tokens. Three question types: noul (yes/no probability), choice (per-label probabilities) and score (ordered scale), several per request. It is built on a larger open-weights model with Liquid-specific customizations and trained for calibration, so a 90% answer should be right at least 90% of the time.

Benchmarks

Artificial Analysis Decision Index v1.0: 65, ranked #1, at least 8 points ahead. AA-Omniscience-Decision non-hallucination rate 85% vs 40% for Microsoft-Decision-1, though more abstention lowers accuracy to 32% vs d1's 41%. Liquid also reports it beats Jev 1.14 across its 112-eval suite (vendor).

Getting started

Call POST /decisions/v1/systemone on the Liquid API with model d1-ultra, or use liquid/d1-ultra on Vercel AI Gateway and OpenRouter. 32,768-token context; $0.042 per M input tokens, output free, each question billed separately. Use it as an agent router, cascade gate or moderation guardrail.