Liquid AI Unveils d1-ultra Decision Model Ranking 1 on AI Benchmark

A New Kind of AI Model

Liquid AI has released d1-ultra, a decision model that works differently from most AI systems. Instead of writing out answers word by word, it reads text, images, or both and responds with a probability score for each possible answer. It does this in a single step and generates zero output tokens.

This matters because developers no longer have to pay for generated text. The model is designed for tasks like sorting messages, checking whether content follows the rules, rating responses, and directing questions to the right system.

How Decision Models Work

A decision model answers focused questions with probabilities rather than free-form text. According to Liquid AI documentation, there are three question types: yes or no questions that return a number between 0 and 1, multiple-choice questions that assign a probability to each label, and scoring questions that place an answer on an ordered scale ¹.

These models are recommended for routing, moderation, guardrails, triage, and replacing language models used as judges. Language models remain better suited for summarization, chat, and code generation.

Performance on Benchmarks

Liquid AI tested d1-ultra across 112 evaluations and says it leads all eight categories in its suite ¹. On the Artificial Analysis Decision Index, which combines four separate tests covering knowledge, reasoning, tool calling, and safety classification, d1-ultra scored 65 and ranked first according to Liquid AI.

For comparison, TypeSafe AI's Jev 1.13 was ranked first on the same index at a score of 58 when Artificial Analysis launched the benchmark in October, with Microsoft's Decision-1 and Fastino Labs' GLiDE tied at 54 each.

Liquid AI claims d1-ultra uses a larger open-weight backbone than its earlier d1 model, combined with custom training methods. The team says improvements are largest on knowledge-based tasks.

Cost and Availability

d1-ultra is available through the Liquid API for $0.042 per million input tokens, the same rate as Jev. Each question counts as its own prompt, and images are billed based on their size. A full-size 1024 by 1024 image costs the equivalent of 1,536 tokens.

The model runs on a dedicated endpoint and supports up to 8 images and 10,000 image patches per request. Support for Vercel AI Gateway and OpenRouter is coming soon, and Liquid AI is offering a playground for testing ¹.

Full details about the model's size, parameter count, and underlying architecture have not been made public.

Key facts

• d1-ultra scores 65 on the Artificial Analysis Decision Index, ranking first among decision models
• The model reads text and images but generates zero output tokens, billing only for input
• Priced at $0.042 per million input tokens, matching TypeSafe AI's Jev
• Liquid AI reports d1-ultra outperforms Jev 1.14 across all 112 evaluation tasks
• Parameter count and model architecture remain undisclosed

Sources

1. Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index — marktechpost.com