Liquid AI d1-ultra: Decision Index Leader and Test Results

👁 113 reads

Liquid AI released d1-ultra on October 15, a decision model that selects an answer from a list and assigns a probability to each option rather than writing text. It scores 65.2 in Artificial Analysis’s Decision Index, making it the new leader among decision models, well ahead of TypeSafe’s Jev 1.13 at 57.5. I checked the AA leaderboard, the manufacturer’s website, pricing and whether the model worked from Poland. On October 19, I also ran it through my set of Polish-language customer-service requests.

What is d1-ultra?

d1-ultra is Liquid AI’s largest decision model. It can process text, images or both together, and answer three types of questions: yes or no, choosing from a list, and rating on a scale. The manufacturer highlights use cases such as detecting a defect in a circuit-board photo or reviewing a customer claim with an attached image. The context window is 32,768 tokens.

Liquid says the model is based on a larger open model with its own modifications, but does not identify the base model or disclose the number of parameters. d1-ultra itself is available only through an API. Only smaller models in the family, d1-3B and d1-omni-600M, have open weights, under the LFM Open License. I covered the earlier, smaller d1 model when the d1 series launched.

d1-ultra’s Decision Index score

Its lead comes mainly from three of the index’s four tests. d1-ultra has the best scores among decision models in these areas:

• When2Call (whether to answer, call a tool, ask a follow-up question or refuse): 92.6, versus 83.7 for the previous leaders, Celeris-1 Decision and GLiDE.
• WildGuardTest (harmful requests and responses): 95.4, versus 95.1 for GLiDE.
• Admitting uncertainty on knowledge questions: 84.5, versus 39.8 for Microsoft-Decision-1.

It is weaker on general knowledge and reasoning. Its accuracy on knowledge questions is 31.6 (the best d1 score is 41.0), and it scores 15.0 on HLE-MultiChoice (Jev 1.13 scores 20.9). General-purpose language models still score higher in this index: Claude Haiku 5.5 in Max mode scores 68.4, but costs about $2,073 per million tasks. Liquid also says d1-ultra beat Jev 1.14 in all 112 evaluations it ran. That is the manufacturer’s result and has not yet been independently replicated.

In the other decision-model leaderboard, JevBench, d1-ultra scores 80.5 in Capability Score and ranks second among API services, behind kettei-fat at 86.3. The open Polish model basal-omni-1.0-ultra scores 82.5 there, but was not measured in the Decision Index, so the two models cannot be compared in one AA table.

Pricing and availability

• Pricing: $0.042 per million input tokens. Output is free because the model does not generate tokens. Liquid counts a 1024 x 1024 px image as 1,536 tokens.
• Where to access it: Liquid’s API (console.liquid.ai), OpenRouter as liquid/d1-ultra using the same /v1/systemone format as other decision models, and Vercel AI Gateway.
• Access from Poland: It worked through OpenRouter without a separate Liquid account when I checked on October 19, 2026. Check with the provider for availability in your country.

My Polish-language test

I used the same set as for the launch of the leaderboard: 50 customer-service requests for department and urgency classification, plus 10 tool-use decisions. I set the answer key on October 13. The test was conducted in Polish; the examples and results below describe that Polish-language test, not a new English-language test.

• Department: 44 out of 47, matching the best decision models on October 13 (Microsoft-Decision-1, Decider, Mercury and Clef). Jev 1.13 scored 41, and basal-omni-1.0-mini scored 42.
• Confidence: At a confidence threshold of 0.9, it got 40 out of 42 right.
• Urgency: 43 out of 50. At a threshold of 0.5, it incorrectly marked 7 cases that could wait as urgent, so you should set a higher threshold using your own data.
• Tool decisions: 10 out of 10.
• Time and cost: Median latency was 351 ms from Poland, including network time. The full test cost $0.00075, or about $12.5 per million tasks of this kind.

I have collected the decision models, both leaderboards and my tests in the AI decision models guide.

// read moreLiquid AI d1: Open Decision Model, AA Benchmark, and PerformanceNewsDecision Index: Jev 1.13 Tops Benchmark for Decision AIBenchmarksDecision Index v1.0 Released by Artificial AnalysisNews

Kinetyra