Liquid AI's d1-ultra Tops Decision Index at 65 for $0.042 Per Million Tokens

Liquid AI's d1-ultra scored 65 on the Artificial Analysis Decision Index, claiming the #1 rank among decision models (a class of AI systems that answer bounded questions with probabilities instead of generated text). The score, reported by Liquid, puts it 7 points ahead of TypeSafe's Jev 1.13, which held first place at 58 when Artificial Analysis launched the index on October 13. Microsoft-Decision-1 and Fastino's GLiDE trailed at 54 each.

The model is available now on the Liquid API at $0.042 per million input tokens — the same rate TypeSafe charges for Jev, according to LiteLLM's documentation. d1-ultra matches Jev's price while, per Liquid's announcement, outperforming it across all tasks in a 112-evaluation comparison suite.

What does a decision model actually do?

d1-ultra does not write text. It reads text, images, or both, and answers typed questions with a probability for each possible answer. It does this in a single forward pass and generates no output tokens — the API reference shows output_tokens always at 0.

Liquid's decision model documentation defines three question types:

• Noul: a yes/no question that returns a probability between 0 and 1.
• Choice: one label from a set, with a probability per label.
• Score: a probability-weighted position on an ordered rubric.

The docs recommend decision models for routing, moderation, guardrails, triage, and replacing LLM-as-judge setups (using a language model to evaluate other models' outputs). They recommend a standard language model for summarization, chat, and code generation.

The economics are the point. Developers running classification, routing, and scoring calls pay only for input tokens, never for generated text. At $0.042 per million input tokens, that undercuts Microsoft-Decision-1 at $31 per million tasks and GLiDE at $119 per million tasks, both as measured by Artificial Analysis.

How does d1-ultra work?

Each request carries a state (the context), optional images, and one or more typed questions. The model evaluates every question against the same input in a single call.

The Liquid AI team states that d1-ultra uses a larger open-weight backbone than its predecessor d1, along with what it describes as heavy Liquid-specific customizations. Combined with Liquid's decision training pipeline, the team says the largest gains come on knowledge tasks.

Liquid built d1-ultra for messy, multimodal cases. Its published examples include spotting a circuit board defect against a spec, and checking a damage claim against a customer's photograph — tasks where the model must reconcile visual evidence with written criteria.

How does it perform on benchmarks?

Liquid compared d1-ultra against its earlier d1 and TypeSafe's Jev 1.14 across 112 evaluations. According to Liquid's announcement, d1-ultra outperforms Jev 1.14 across all tasks at the same price. It also leads all 8 categories in Liquid's own suite — a claim that, like the headline index score, remains vendor-reported.

The Artificial Analysis Decision Index averages four evaluations, weighted at 25% each:

• AA-Omniscience-Decision: knowledge and hallucination
• HLE-MultiChoice: reasoning
• When2Call: tool calling
• WildGuardTest: safety classification

For context, Artificial Analysis's October 13 launch article ranked Jev 1.13 first at 58, with GLiDE and Microsoft-Decision-1 tied at 54. Liquid claims d1-ultra's 65 takes the top spot, and separately claims the lowest calibration error in its tests — a property that matters for developers setting routing thresholds on probability outputs.

How much do the unknowns matter?

Several core facts remain undisclosed. Liquid has not published d1-ultra's parameter count, its context length, or a per-benchmark breakdown of the 112 evaluations. Access is API-only; no downloadable weights or quantized variants have been announced.

The comparison table tells the story:

• d1-ultra (Liquid AI): text and image input, API-only, AA Decision Index 65 (Liquid-reported), $0.042 per 1M input tokens, no output-token billing.
• d1 (Liquid AI): text and image input, API-only, index score and price not disclosed.
• Jev (TypeSafe AI): text-only input, closed weights, index score 58 (v1.13, October 13), $0.042 per 1M input tokens.
• Microsoft-Decision-1 (Microsoft): parameters, input types, and weights not disclosed, index score 54 (October 13), $31 per 1M tasks (AA).
• GLiDE (Fastino Labs): parameters, input types, and weights not disclosed, index score 54 (October 13), $119 per 1M tasks (AA).

Every number above 65 in the leaderboard, in other words, traces back to Liquid's own reporting pending independent verification.

How do you deploy it, and what does it cost in practice?

d1-ultra runs on the Liquid API through the /decisions/v1/systemone endpoint with "model": "d1-ultra". Liquid says Vercel AI Gateway and OpenRouter support is coming soon. Developers can test the model first in the d1 Playground.

Billing covers input tokens only, but image pricing carries a multiplier worth calculating. Per Liquid's pricing notes, images cost 1.5 tokens per 32×32-pixel patch. A 1024×1024 image therefore costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images — so 3 questions on 1 image bill that image 3 times.

The endpoint documentation caps requests at 8 images and 10,000 image patches, with the full JSON body limited to 4.5 MB.

Why this matters

Decision models attack a specific cost center in AI infrastructure: classification, routing, moderation, and judging workloads currently run on general-purpose LLMs that generate billed output tokens for questions with bounded answers. If Liquid's vendor-reported numbers hold up under independent testing, a $0.042-per-million-token multimodal model topping the Decision Index would pressure rivals whose per-task pricing runs hundreds of times higher. Until Liquid discloses parameters, weights, or third-party-verified benchmarks, buyers should treat the 65 as a claim — one they can now test themselves in the d1 Playground.

By Marcus Bennett