Editorial illustration for Liquid AI releases multimodal d1-ultra, scores 65 on AA Decision Index
Liquid AI's d1-ultra Tops Decision Index Benchmark
Liquid AI releases multimodal d1-ultra, scores 65 on AA Decision Index
Liquid AI put a number on its newest model Tuesday: 65 on the Artificial Analysis Decision Index, enough for the top rank on that benchmark, according to the company. The model, called d1-ultra, is built on a larger open-weight backbone than its predecessor d1, though Liquid hasn't disclosed parameter count or context length. It's live now through the Liquid API at $0.042 per million input tokens, with no downloadable weights or quantized versions on offer.
What makes d1-ultra different from a typical chatbot is what it doesn't do. It takes text, images, or both, and instead of generating output tokens, it returns a probability for each possible answer to a typed question, all in a single forward pass. That matters for developers running classification, routing, or scoring tasks who don't want to pay for generated text they're going to throw away anyway. Liquid says the model leads all eight categories in its internal 112-evaluation suite, though that breakdown, like the Decision Index score, comes from the company itself rather than an independent source.
So what exactly counts as a "decision model," and how does one answer a question without writing a word of it?
Liquid AI compared d1-ultra against its earlier d1 and TypeSafe’s Jev 1.14 across 112 evaluations. According to Liquid AI’s announcement, d1-ultra outperforms Jev 1.14 across all tasks at the same price.
Why this matters
A model that scores probabilities instead of generating tokens is a narrow bet, but it's the right bet for a huge chunk of production AI work: routing, moderation, classification, scoring calls that don't need prose, just a confident answer. At $0.042 per million input tokens and no output tokens to pay for, d1-ultra is pitching itself directly at teams burning money on chat-style models to do what's really a lookup-and-score task. The multimodal piece, handling images alongside text in one forward pass, extends that pitch to content moderation and document triage without bolting on a separate vision model.
We'd treat the 112-evaluation comparison and the #1 Decision Index ranking as Liquid AI's framing, not independent verification, since the numbers come from their own announcement. Still, the architecture choice itself is worth watching: if decision-only models keep beating general-purpose LLMs on cost and latency for classification work, expect more vendors to split "answer generation" from "decision making" as separate product lines.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Feature Elicitation for SAE Steering Reveals Deceptive Defanging in Small LLMs
From Archives SSR Reformulates Reasoning as Selection for Multimodal Search Agents
Common Questions Answered
What score did Liquid AI's d1-ultra achieve on the Artificial Analysis Decision Index?
Liquid AI's d1-ultra scored 65 on the Artificial Analysis Decision Index, which according to the company is enough for the top rank on that benchmark. This score demonstrates the model's superior performance compared to other decision-making models in its category.
How does d1-ultra's pricing compare to other models like TypeSafe's Jev 1.14?
D1-ultra is priced at $0.042 per million input tokens with no output token charges, making it cost-effective for production AI work. According to Liquid AI's testing across 112 evaluations, d1-ultra outperforms Jev 1.14 across all tasks at the same price point.
What are the key differences between Liquid AI's d1-ultra and its predecessor d1?
D1-ultra is built on a larger open-weight backbone than its predecessor d1, though Liquid AI has not disclosed the specific parameter count or context length. The upgraded model represents a significant improvement in decision-making capabilities as evidenced by its higher benchmark score.
What types of production AI tasks is d1-ultra designed to handle?
D1-ultra is specifically designed for tasks like routing, moderation, classification, and scoring decisions that require confident answers rather than prose generation. The model scores probabilities instead of generating tokens, making it ideal for teams currently using chat-style models for lookup-and-score operations.
Is d1-ultra available for download or does it only run through the Liquid API?
D1-ultra is only available through the Liquid API and does not offer downloadable weights or quantized versions. The model went live on the Liquid API immediately after its announcement, providing access exclusively through their cloud service.
