Liquid AI d1-Ultra: A $0.042 Decision Model Claiming the Top Spot

Liquid AI has released d1-ultra, a decision model that reads text, images or both, and answers closed questions with a probability instead of a paragraph. According to Liquid's launch post, it costs $0.042 per million input tokens, is available on the Liquid API today, and "outperforms all other decision models" on the Artificial Analysis Decision Index "by at least 8 points."

That last sentence is the claim that will travel. This post lays out what d1-ultra is, what the numbers say, and what is and is not checked. If the category is new to you, start with our practitioner guide to decision models.

TL;DR: questions readers will ask

What is a decision model, briefly?

A decision model does not write text. You give it a situation and a closed set of answers, and it returns a probability for each in a single pass. You pay only for input, and you never have to parse a malformed reply. That makes it a drop-in replacement for many classification and triage jobs that people currently hand to a general chat model.

Liquid's examples are deliberately physical and messy: spotting a defect on a circuit board against its specification, or checking a damage claim against a customer's photo. The request in the post's code sample sends a base64 image plus a state description and one noul question asking whether the board has a defect, and reads the answer from answers.defect.noul.

What is new compared with d1?

Liquid says d1-ultra is based on a larger open-weight model than d1, with "a series of Liquid-specific heavy customizations," and that the gains show up especially on knowledge-related tasks. It also says it trained the model specifically for calibration, because "most decision models are overconfident." That matters because a probability is only useful if 0.9 means about nine in ten.

We previously covered the earlier version when it learned to read pictures, in Liquid AI d1 now sees images. The pricing story is similar: the same $0.042 per million input tokens that Jev charges, now with a larger model behind it.

Claimed versus verified

The caveat is the usual one for vendor launches. A lab that publishes a benchmark of its own model against competitors controls the framing: which evaluations, which versions, which prompts. Eight points on an index that did not exist a month ago is interesting, but the index itself is young. For context, our ranking of the top 10 decision models shows how quickly the leader has changed in the few weeks since the category appeared, including open-weight models from Perplexity and others that run on your own hardware.

How to check it yourself

1. Build a labeled set from your own data. Even 200 to 500 examples with known answers will tell you more than any vendor table.
2. Measure accuracy and calibration separately. Bucket predictions by stated probability and check whether the 0.8 bucket is right about 80 percent of the time.
3. Compare on cost per decision, not just per token. Each question is billed as its own prompt including every image, so a request with five questions on one photo costs five times the image tokens.
4. Run the same set against Jev 1.14 and one open-weight alternative. Liquid's claim is relative, so test it relatively.
5. Watch for the OpenRouter listing. Liquid says the model will arrive there soon, which will make side-by-side tests simpler.

A minimal request, adapted from Liquid's own example, looks like this:

import base64, os, requests

image = base64.b64encode(open("board.jpg", "rb").read()).decode()
response = requests.post(
    "https://api.liquid.ai/decisions/v1/systemone",
    headers={"Authorization": f"Bearer {os.environ['LIQUID_API_KEY']}"},
    json={
        "model": "d1-ultra",
        "images": [f"data:image/jpeg;base64,{image}"],
        "state": "Camera image of a circuit board on the production line.",
        "questions": {"defect": {"type": "noul", "instructions": "Does this circuit board have a defect?"}},
    },
)
print(response.json()["answers"]["defect"]["noul"])

Where does it fit among other decision tools?

Decision models are cheap enough that the interesting question is where they sit in a pipeline. A common pattern is to use one as a first-pass gate: route claims, tickets, defects or content flags through a probability, send only the uncertain middle band to a larger reasoning model or a human, and log the rest. Calibration is what makes that work, which is why Liquid leads with it. A model that is accurate but overconfident would push too many wrong answers past the gate.

For another example of the "small specialist beats a large generalist" argument, see our write-up of AI21 and Deel's payroll results, where a vendor also published cost and accuracy numbers that readers should check against their own data.

A short timeline of the category

The decision-model category is only weeks old, which is part of why a single launch can claim a lead so quickly.

• September 15, 2026: TypeSafe AI launched Jev, a hosted model that returns a typed answer with a probability, at $0.042 per million input tokens, according to our top 10 ranking.
• Early October: Liquid shipped the first d1 and then a version that reads images, and Perplexity published pplx-decider v1.1 with a reported Decision Index 0.3 score of 61.56 against Jev's 57.9, per the same ranking. Perplexity's number comes from its own model card.
• October 15, 2026: Liquid launches d1-ultra and says it leads the Artificial Analysis Decision Index by at least 8 points.

Two things stand out. First, every claimed lead rests on an index that has been around for a very short time and that vendors report on themselves. Second, the price has barely moved: Liquid matched Jev's $0.042 rather than undercutting it, so the competition at the top is now about accuracy and calibration, not cost. Perplexity's reported $0.02 per million input tokens remains the lowest hosted price we have seen mentioned, though we could not confirm it on the model card.

What would make the claim more convincing?

Liquid's post is short on detail, and several pieces of information would help a buyer judge it.

• The list of 112 evaluations. Knowing which tasks were included, and how many were text versus image, shows whether the average hides weak spots.
• Calibration numbers. The post says calibration was a training goal but gives no figure here such as expected calibration error. For a probability product, that is the metric that matters most.
• The base model. Liquid says it built on a larger open-weight model. Which one, and under what licence, affects how portable the approach is and whether the weights could be released.
• Latency. Earlier Liquid models highlighted edge latency. Nothing in this post says how fast d1-ultra answers a single question.
• Rate limits and quotas. The post points to the API but not to production limits.

Until those appear, the sensible reading is that d1-ultra is a credible new entry from a lab with a track record in the category, and that its lead is plausible but unproven outside Liquid's own evaluations.

What this means for what you build or pay

At $0.042 per million input tokens, a 1,536-token image question costs about $0.00006. A million such questions would cost roughly $65 on those numbers, before any retries. If Liquid's accuracy claim holds for your task, the savings against a general frontier model are large. If it does not, the low price will not matter, because a wrong decision at scale is expensive. The practical move is to run a small paid trial rather than rely on the leaderboard.

Related reading

• What are decision models? A practitioner guide
• Liquid AI d1 now sees images
• Top 10 decision models in October 2026
• Jev 1.14: claimed vs verified
• AI21 and Deel payroll: claimed vs verified

Prices, benchmarks and availability are as stated by Liquid AI on October 15, 2026 and may change.