Liquid AI's d1-ultra charges the same $0.042 per million input tokens as TypeSafe's Jev, with no output fee, but it bills every image again for every question you ask about it. Its first place on the Artificial Analysis Decision Index was earned on four tests that contain no images. If you run claims checks, defect inspection or policy classification on a decision model, the switch could be worth it. Re-run your own decision set and price it per question before you move any traffic.
Liquid released d1-ultra on October 15, 2026, two days after Artificial Analysis published Decision Index v1.0 on October 13. The launch post calls it "the world's best decision model." What follows is what the evidence supports, what it costs once photos are involved, and what to test.
What d1-ultra Is and What It Costs
A decision model is a model that reads an input once and returns a probability for each option in a fixed set of answers, instead of generating text. d1-ultra does this for text, images or both. Per Liquid's announcement, it answers three question types (yes/no, choice from a list of labels, and a score on a scale) in a single forward pass, and Liquid built it on "a larger open-weight model" it does not name, with heavy customization aimed at better-calibrated probabilities.
The price is simple on paper. Liquid's pricing page lists d1-ultra at $0.042 per million input tokens, $0.042 per million cached input tokens, and $0.00 for output. The original d1 is $0.02 per million input tokens, so the new model costs 2.1 times its predecessor. New organizations get $2 in free credit on Liquid Console.
That $0.042 rate is the same one Jev has charged since at least September, which is why Liquid can claim "higher accuracy at the same price." Vercel added d1-ultra to AI Gateway on launch day with a 32,768-token context window, no markup and no platform fee, including on bring-your-own-key requests. Gateway logs decision calls and counts them against budgets alongside your other model traffic. The model is hosted only: Liquid's docs list d1 and d1-ultra on the Liquid API, OpenRouter and Vercel AI Gateway, while the smaller d1-3B and d1-omni-600M went up on Hugging Face as open weights. Jev already has a tool page here, and d1-ultra is reachable through Vercel.
Does d1-ultra Really Lead the Decision Index?
Yes, on the live board, by a little under or a little over eight points depending on which Jev you compare. The v1.0 launch article, dated October 13, predates d1-ultra and ranks Jev 1.13 first among decision models at 58, with GLiDE and Microsoft-Decision-1 tied at 54. One language model, Claude Haiku 5.5 (Low), scores 61 and sits above every decision model in that article; GPT-6 Luna (Low) scores 55.
The live index page has since added d1-ultra and Jev 1.14. BenchLM's mirror of Artificial Analysis's decision-model rows, in its October 19 snapshot, shows d1-ultra at 65.2, Jev 1.13 at 57.5 and Jev 1.14 at 57.1. Liquid's claim of a lead "by at least 8 points" holds against Jev 1.14 and rounds up against Jev 1.13. Crypto Briefing reported the same 65 against 57 for Jev 1.14 and added that the model "pairs a high non-hallucination rate with lower accuracy," so the lead comes mostly from abstaining more often. Read the per-test rows before you treat 65 as a sign that d1-ultra gets more answers right.
The index itself is worth understanding before you lean on it. It is a weighted average of four tests at 25% each: AA-Omniscience-Decision (500 private questions with an "I don't know" option, split evenly between accuracy and non-hallucination rate), 513 multiple-choice questions from Humanity's Last Exam, NVIDIA's When2Call tool-calling test, and Ai2's WildGuardTest safety classifier set, per Artificial Analysis. None of those four is an image task. d1-ultra's pitch is multimodal, so the leaderboard number tells you nothing about how it reads a damaged bumper.
The comparison Liquid ran is its own: 112 unique evaluations against d1 and Jev 1.14, with the claim that d1-ultra wins "across all tasks." The scores are in chart images on the launch post rather than in text, and OrcaRouter points out that HLE-MultiChoice, the index's hardest reasoning component, is not among the results Liquid reports.
Where the Image Bill Comes From
Every question re-bills every image in the request. Liquid's documentation is explicit: each question is billed for its own text plus all images supplied, and a 1024 by 1024 image costs 1,536 image tokens per question, so two questions about one photo cost 3,072 image tokens before any text. Image tokens are computed as 1.5 tokens per 32 by 32 pixel patch.
Jev's model is different. TypeSafe bills the state (the data being judged) once per request, not once per question, which makes batching several questions into one call cheaper than separate calls. With d1-ultra, batching saves you request overhead but not image tokens.
Here is what that does at volume, using Liquid's formula and its $0.042 rate, image tokens only:
| Workload, per 1 million inputs | Image tokens per input | d1-ultra image cost |
|---|---|---|
| One 1024x1024 photo, 1 question | 1,536 | $64.51 |
| One 1024x1024 photo, 10 questions | 15,360 | $645.12 |
| One 512x512 photo, 10 questions | 3,840 | $161.28 |
| One 1024x1024 photo, 10 questions, on d1 at $0.02 | 15,360 | $307.20 |
Steel-man the other side: these are small numbers. $645 per million claims is noise next to an adjuster's time, and Artificial Analysis puts Claude Haiku 5.5 (Low) at $184 per million text tasks against $42 for Jev 1.13. The cost lesson is about design. A claims flow that asks twenty narrow questions per photo, which is how the decomposition pattern we covered for Jev works best, pays twenty times for the picture. Resolution is the other lever: halving each side cuts image tokens by four.
The documented limits will shape your pipeline too. Requests take at most 8 images and 10,000 patches in total, the full JSON body must stay under 4.5 MB, and images must be base64-encoded because remote URLs are not accepted. A multi-angle vehicle claim or a circuit board shot at full resolution can hit those caps before it hits your budget.
What Liquid Has Not Shown Yet
The case for d1-ultra rests on calibration, and calibration is the claim with the least public evidence. Artificial Analysis found that every decision model it tested was overconfident, with accuracy below the stated probability "for all decision models" at almost every level, worst on knowledge and reasoning tasks. Liquid says it trained d1-ultra specifically against that. Its predecessor scored 41 on Omniscience-Decision accuracy, the best among decision models, but 2 on non-hallucination, meaning it almost never chose "I don't know." That is the failure a calibrated successor has to fix, and the public record does not yet show it fixed.
Three gaps are concrete. Liquid publishes no image benchmark that a third party has run. Runtime Wire's launch report attributes the 112-test result to Liquid alone. And the API surface is still settling: Vercel's own changelog names the same yes/no question type three ways (boolean in the AI SDK, predicate in its Decisions API, noul in the TypeSafe-compatible client), so pin your client version and test the type names before you wire it into production.
If you have already compared decision models, the pattern is familiar. Microsoft shipped Decision-1 at Jev's price with rankings on tests it did not name, and Clef beat Jev at routing while losing at judgment. Vendor leads in this category have so far been narrow and task-specific.
What to Do Before You Switch
Run d1-ultra in shadow on your own labeled decisions, priced per question. The work fits in a month.
This Week:
- Pull 500 labeled cases from your current decision workload, including the hard and ambiguous ones, and run them through d1-ultra alongside your incumbent. The $2 Liquid credit covers a small first pass.
- Count questions per input in your production flow and multiply your image token count by it. If you ask more than three or four questions per photo, price the image downscaled to 512 pixels and check that accuracy holds.
- Bin d1-ultra's probabilities (0.9 to 1.0, 0.8 to 0.9, and so on) and compare each bin's stated confidence with its actual accuracy. That reliability check is the claim you are paying for.
This Month:
- Set your routing thresholds from the reliability table, not from the vendor's chart. Send anything below the threshold to a human or a larger model, as the Jev vs Clef vs Haiku buyer's guide describes.
- Read d1-ultra's per-test scores on the live index, especially Omniscience-Decision accuracy against non-hallucination and HLE-MultiChoice, which Liquid's post does not report.
- If you route through OpenRouter or Vercel, confirm the listed price and context window match Liquid's direct API before you commit volume.
The Bottom Line
Decision models are following the path text LLMs took in 2023 and 2024: a new release claims the top of a fresh leaderboard within days, at a price matched to the incumbent, and the buyer's own evals are the only ranking that settles anything. d1-ultra adds images to that category, which is useful for claims and inspection work, and it adds a billing rule that punishes many-question designs. If you already use OpenAI's Decisions API or Jev, the question is narrow: on your images, at your question count, is d1-ultra better calibrated? Your 500-case shadow run answers it.
Price the photo per question first, then read the reliability table.
Continue Reading
- TypeSafe's Jev Scores 62.6% Asked Once and 95% Split Five Ways
- Microsoft Ships Decision-1 at Jev's Price, Ranked on Unnamed Tests
- OpenAI's Decisions API Skips Output Fees and the Cache Discount
- Jev vs Clef vs Haiku: Decision Models Cut Cost 25x, Not Errors
- Cloudflare's Clef Beats Jev at Routing and Loses at Judgment
- Apple's On-Device Model Is 99% Sure. Sample It Five Times.