Liquid AI released d1-ultra on October 15, 2026, at $0.042 per million input tokens — the same rate Microsoft charges for Microsoft-Decision-1. Output is free. On Artificial Analysis’s Decision Index, the independent leaderboard for this class of model, d1-ultra sits first at 65.2 among 13 decision models. Liquid calls it the “World’s Best Decision Model.” The gap to second place is 7.7 points, not the “at least 8” Liquid advertised.
Here’s the twist that makes the score interesting: d1-ultra’s knowledge-accuracy subscore is lower than Liquid’s own cheaper d1. It abstains more. That is the product story, not a bug in the table.
Liquid d1-ultra price: $0.042 in, $0.00 out
Liquid’s pricing page, checked October 18, 2026, lists d1-ultra at $0.042 per million input tokens, $0.042 for cached input and $0.00 for output. The older d1 sits at $0.02 across those same three lines. Liquid’s launch blog post quotes the $0.042 figure and says the model is on Liquid’s API today.
| Model (Liquid pricing, Oct 18) | Input / 1M tokens | Cached input | Output |
|---|---|---|---|
| d1 | $0.02 | $0.02 | $0.00 |
| d1-ultra | $0.042 | $0.042 | $0.00 |
That $0.042 line matches Microsoft-Decision-1’s public rate. Both charge for input and waive output, which fits a model that returns probabilities instead of generating text. Liquid’s blog said d1-ultra would land “soon” on Vercel and OpenRouter; systemonemodels.org, which tracks System One / decision-model listings, said both gateways had listed it by October 16. Liquid’s own docs do not publish a context-window figure for d1-ultra; OpenRouter and Vercel listings cite 32,768 tokens — attribute those to the gateways, not to Liquid’s blog.
Images count as input. Liquid bills them at 1.5 tokens per 32×32 patch, so a 1024×1024 image is 1,536 tokens. Each question is billed for its text plus all images attached to it. The endpoint example in Liquid’s post is POST https://api.liquid.ai/decisions/v1/systemone with model name d1-ultra.
What a decision model is, and what d1-ultra returns
d1-ultra is multimodal: it takes text and images and returns probabilities. It generates no tokens. Question types match the earlier d1 — Liquid’s names are noul (yes/no probability), Choice and Score. Liquid’s use-case sketches include checking a circuit board against a spec drawing and scoring a damage claim against a photo. Those are Liquid’s examples, not independent case studies.
Liquid says the model is built on a larger open-weight base with Liquid customizations. It does not name the base model or publish a parameter count. systemonemodels.org notes there are no downloadable weights: commercial hosted API only. Liquid’s d1-family docs (for the earlier d1) list limits such as up to eight images per request; treat those as the family defaults unless Liquid publishes d1-ultra-specific caps.
Decision Index 65.2: the lead, and the catch on “8 points”
On the Artificial Analysis Decision Index snapshot mirrored at BenchLM for October 15, 2026, d1-ultra leads at 65.2. The next names in that snapshot: Jev 1.13 at 57.5, Jev 1.14 at 57.1, Microsoft-Decision-1 fifth at 53.8, Liquid’s own d1 at 50.3 and GPT-6 Luna Decisions at 48.5. Liquid said the model “outperforms all other decision models [on the] Artificial Analysis Decision Index by at least 8 points.” Against Jev 1.13, the published gap is 65.2 − 57.5 = 7.7. systemonemodels.org flags that discrepancy; we do too. Say Liquid claimed “at least 8”; the published #2 gap is 7.7.
| Decision Index (AA / BenchLM, Oct 15 snapshot) | Score |
|---|---|
| Liquid d1-ultra | 65.2 |
| Jev 1.13 | 57.5 |
| Jev 1.14 | 57.1 |
| Microsoft-Decision-1 | 53.8 |
| Liquid d1 | 50.3 |
| GPT-6 Luna Decisions | 48.5 |
d1-ultra’s published subscores on that snapshot: knowledge accuracy 31.6, non-hallucination 84.5, HLE-MultiChoice 15.0, When2Call 92.6, WildGuardTest 95.4; cost about $31.8 per 1M tasks; latency 93 ms. Compare knowledge and abstention with Liquid’s d1 on the same board: d1’s knowledge accuracy is 41.0, but its non-hallucination score is 1.7. d1-ultra knows less on that subscore and refuses more. That is how you get a higher composite while looking weaker on raw knowledge.
Liquid also publishes its own 112-eval suite versus d1 and Jev 1.14. Those numbers are vendor-run. Prefer the independent Decision Index for ranking claims; treat Liquid’s suite as Liquid’s.
Same price as Microsoft-Decision-1, different scoreboard place
We covered Microsoft-Decision-1 when it launched at the same $0.042 input rate with free output. On this Decision Index snapshot it sits fifth at 53.8 — more than 11 points behind d1-ultra. Price alone does not pick a winner. If your workload is “pick from a list and tell me how sure you are,” the abstention behavior and the independent composite matter more than a vendor’s “world’s best” line.
What is still open: whether Liquid names the base model, whether independent runs reproduce the 65.2, and how OpenRouter/Vercel latency and availability settle once traffic moves off Liquid’s API. We track decision-model pricing and index moves in our LLM News feed.
Frequently asked questions
How much does Liquid d1-ultra cost?
On Liquid’s pricing page, checked October 18, 2026: $0.042 per million input tokens, $0.042 for cached input and $0.00 for output. Images are billed as input at 1.5 tokens per 32×32 patch (a 1024×1024 image is 1,536 tokens).
How does d1-ultra compare with Microsoft-Decision-1?
Both list $0.042 per million input tokens with free output. On the Artificial Analysis Decision Index snapshot for October 15, 2026, d1-ultra leads at 65.2 and Microsoft-Decision-1 is fifth at 53.8. We covered Microsoft’s model separately when it launched.
Is Liquid d1-ultra output free?
Yes. Liquid’s pricing page lists output at $0.00. The model returns probabilities and does not generate tokens, so there is no output-token charge.
Are there open weights for d1-ultra?
No downloadable weights are published. It is a commercial hosted API on Liquid (and listed on gateways such as OpenRouter and Vercel). Liquid says it is built on a larger open-weight base with Liquid customizations but has not named that base or a parameter count.
What is a decision model?
A decision model takes a question (and, for multimodal ones like d1-ultra, images) and returns probabilities over answers you supply — yes/no, a choice among options or a score — instead of writing free-form text. Liquid’s types for d1-ultra are noul, Choice and Score.
What is Liquid d1-ultra’s Decision Index score?
65.2 on the Artificial Analysis Decision Index (BenchLM mirror of the October 15, 2026 snapshot), first among 13 decision models. Liquid claimed an “at least 8 point” lead; the published gap to #2 (Jev 1.13 at 57.5) is 7.7 points.
Sources
All accessed October 18, 2026, between about 8:45 and 9:15 AM ET.
- Liquid AI blog, “d1-ultra” (Oct 15, 2026; launch, price, multimodal decision API, image token billing, endpoint, use cases, “World’s Best” / “at least 8 points” claims, 112-eval suite): liquid.ai/blog/d1-ultra
- Liquid AI pricing (d1 and d1-ultra input / cached / output rates; checked Oct 18, 2026): liquid.ai/pricing
- BenchLM mirror of Artificial Analysis Decision Index (Oct 15, 2026 snapshot: d1-ultra 65.2 and subscores; Jev 1.13/1.14; Microsoft-Decision-1 53.8; d1 50.3; GPT-6 Luna Decisions 48.5): benchlm.ai/benchmarks/artificial-analysis-decision-index
- Artificial Analysis (Decision Index primary): artificialanalysis.ai
- systemonemodels.org, d1-ultra model page (updated Oct 16, 2026; hosted-only, gateway listings, 7.7-point gap note, 32,768 context attributed to gateways): systemonemodels.org/models/d1-ultra
- Liquid docs, d1 family (image limits and question types for the d1 line): docs.liquid.ai/lfm/models/d1
Related: Microsoft-Decision-1 price and speed claims · More LLM news
Last updated: October 18, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
