d1-ultra
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.4/10
Paid self-serve Liquid API with the same docs and TypeSafe-SDK compatibility as d1. Listed on OpenRouter and Vercel AI Gateway on launch day. Liquid does not call it GA and publishes no SLA.
- Capability7.9/10
Noul / Choice / Score, text and images, zero generated tokens. #1 on the independent Artificial Analysis Decision Index v1.0 (65) with 92 ms per task. Calibration is claimed with an internal chart, but no method or ECE figure is public.
- Adoption25/100
Launch post at ~1.4k likes / ~255k views within a day, AA write-up, Vercel and OpenRouter listings on day one. Backed by an established lab. No production cases reported yet.
Vendor claims
- Leads all eight categories of Liquid's own 112-evaluation suite and outperforms Jev 1.14 across all tasks at the same price[Vendor claim — not independently verified]
- Lowest calibration error of any model Liquid tested: when d1-ultra says 90%, it is right at least 90% of the time[Vendor claim — not independently verified]
- Supports 44 languages[Vendor claim — not independently verified]
News covering this model
d1-ultra is the second hosted decision model from Liquid AI, released on 15 October 2026, about two weeks after d1. Liquid pitches it at harder, messier inputs: checking a circuit board against its spec, or a damage claim against the customer’s photo.
It uses the same contract as d1 and Jev. You send a state (text, images or both) plus named typed questions, and get a probability for every answer from one forward pass. Liquid’s docs list it next to d1 on the same endpoint; only the model field changes.
Specs
| Attribute | Value |
|---|---|
| Company | Liquid AI |
| Launch | 15 Oct 2026 (blog and X, 14:00 BRT) |
| Status | Paid self-serve API. Liquid does not call it GA |
| Endpoint | POST https://api.liquid.ai/decisions/v1/systemone with "model": "d1-ultra" |
| Question types | noul, choice, score |
| Input | Text and images (images are billed as input tokens: 1.5 tokens per 32×32 patch, so a 1024×1024 image is 1,536 tokens) |
| Context | 32,768 tokens |
| Pricing | $0.042 / M input, no output tokens. Each question is billed as its own prompt, images included |
| Distribution | Liquid API, Vercel AI Gateway (liquid/d1-ultra) and OpenRouter, all on 15 Oct |
| Weights / base model | Closed. Liquid says only that it is “based on a larger open-weight model with a series of Liquid-specific heavy customizations” |
Artificial Analysis Decision Index v1.0
On 16 October Artificial Analysis posted its numbers for d1-ultra. This is an independent board, separate from Hugging Face’s Decision Index 0.3 and from JevBench, so the scores don’t compare across boards.
| Metric (AA, text of the post) | d1-ultra | Reference |
|---|---|---|
| AA Decision Index v1.0 | 65 (#1) | Jev 1.14: 57 · GLiDE and Microsoft-Decision-1: 54 |
| AA-Omniscience-Decision non-hallucination rate | 85% | d1: 2% |
| AA-Omniscience-Decision accuracy | 32% | d1: 41% |
| Time per task | 92 ms (3rd lowest) | Celeris-1 Decision 63 ms · Jev 1.14 78 ms |
| Cost per 1M tasks | $32 | Microsoft-Decision-1 $31 · d1 and Mercury Decide $15 · GPT-6 Luna (Decisions) $79 |
The lead comes mostly from one behaviour. d1-ultra answers “I don’t know” when it lacks the knowledge, where d1 almost never did. Its raw accuracy on that knowledge test went down, not up. That is a useful property for automation with thresholds, but it is not the same as “smarter on every task”.
Checked on 16 Oct 2026: Liquid’s pricing page lists d1 at $0.02 / M input and d1-ultra at $0.042 / M, and OpenRouter now lists d1 at $0.02 / M too. d1 launched on OpenRouter at $0.04 / M on 1 Oct, and Liquid’s own 5 Oct benchmark note still used $0.04, so the price was cut after that. Our d1 page is updated to $0.02 / M, which matches AA.
Liquid’s own numbers (UnverifiedClaim)
Liquid compares d1-ultra with d1 and Jev 1.14 on 112 evaluations of its own and says it wins every category, with better calibration. The charts are Liquid’s. The suite, items and calibration method are not published, so none of this enters our scores beyond the “claimed” band.
Fit / anti-fit
Fit when you need: a hosted decision model that knows when to abstain, image checks against a written spec, or a drop-in A/B against Jev or d1 with the same request shape.
Anti-fit when you need: local or on-prem inference (no weights), long inputs (32K context, against 262K on Jev 1.14), an SLA, or the cheapest option (d1 costs less than half as much per task on AA).
Limits
- No public SLA, no published ECE or calibration method.
- Base model, size and training data are not disclosed.
- Not on Hugging Face’s Decision Index 0.3 (open models only for the private tasks) or on JevBench.
Why it is in the catalog
Same reasoning as d1: Liquid presents it as its own trained decision model, with a public, documented typed API that returns probabilities. It gets its own page because it is a separate model id with its own price, board position and behaviour.
Score working
- Maturity 5.4 = availability 6 × 0.30 + docs 7 × 0.25 + integrations 5 × 0.25 + support 3 × 0.20 = 1.80 + 1.75 + 1.25 + 0.60. Same as d1.
- Capability 7.9 = types 10 × 0.30 + calibration 5 × 0.25 + latency 8 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.25 + 2.00 + 1.60 = 7.85. Accuracy 8 because the AA board is independent (same rule as GLiDE). Latency 8: AA’s 92 ms per task, no p99.
- Adoption 25: strong launch and two distribution channels on day one, no production reports.