Liquid AI d1-ultra vs Liquid AI d1: Same Endpoint, Same Questions, Twice the Price — Here's the Crossover

The shortest honest answer is this: Liquid AI d1-ultra costs $0.042 per million input tokens and Liquid AI d1 costs $0.02 per million input tokens, both are called at the same URL with the same three question types, and the more expensive one is the one to use when being wrong is more expensive than doubling your per-call rate. That is the entire decision, and it is unusually clean because nothing else about the two models differs structurally. They were shipped ten days apart — d1 on October 5, 2026 as Liquid AI's first decision model, d1-ultra on October 15 as the family's most capable member — and both were built to do the same thing: accept a state of text, images or both plus one or more typed questions, and return probabilities from a single forward pass with zero output tokens generated and nothing to stream. Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What makes the pair worth a piece rather than a footnote is that they are substitutes, not a ladder with a middle rung. You do not run d1 and fall back to d1-ultra on hard cases the way you would a small and a large generative model, because both are gated behind the same model string in the same request body. Switching costs one word in a JSON payload. So the question is not whether the upgrade is technically available — it is whether the extra $0.022 per million input tokens buys anything your particular workload can feel, and that is answerable with the numbers Liquid AI has published.

What actually differs, in one pass

The disclosed differences are fewer than the price gap suggests, and the ones that are missing matter as much as the ones that are present.

• Price — $0.042 per million input tokens for d1-ultra versus $0.02 for d1, both billed on input only with no output charge, both counting images as text at 1.5 tokens per 32×32-pixel patch.

• Backbone — d1-ultra is described as built on "a larger open-weight model with a series of Liquid-specific heavy customizations"; d1's announcement says only that it was built for the first generation of the family, with no backbone named either.

• Language coverage — Liquid AI's launch post for d1-ultra states 44 languages. No comparable figure was published for d1.

• Index placement — d1-ultra is claimed to lead the Artificial Analysis Decision Index by at least eight points; d1 was announced with six real-application comparisons against GPT-6.1 Sol and Claude Opus 5.5 rather than an index score.

• Parameter count — disclosed for neither. d1-ultra's context window is not stated by the vendor either, and appears only in third-party gateway catalogues at 32,768 tokens, which we treat as unconfirmed rather than as a specification.

• Latency — d1 was published at 200 to 300 ms for a text decision. d1-ultra has no published latency figure at all.

That last pair of gaps is the thing to notice. The newer, more expensive model is the one whose spec sheet is emptier. You are being asked to pay 2.1 times the rate for a model whose size, context length and speed are all undisclosed, on the strength of a calibration claim and an index number the vendor itself reported.

The claim Liquid AI is actually making

Read the d1-ultra post closely and the pitch is not accuracy. It is that accuracy alone is insufficient for a decision model, because a decision model also emits probabilities, and most of them are overconfident. A classifier that says 0.97 and means 0.72 will wreck a threshold-based pipeline even if its top-1 accuracy is excellent, because every automated action you take above that threshold is being taken on a false premise. d1-ultra was trained, the post says, specifically so its probabilities are better calibrated — and the release reports that across 112 unique evaluations it beats Jev 1.14 on every task with better calibration at the same price.

This is a genuinely different sales argument from "smarter," and it is also the single hardest claim in either release for an outsider to verify. Nothing in the announcement lets you check calibration on your own data before committing. If your pipeline takes the argmax and ignores the probability, the calibration pitch does not apply to you, and d1-ultra's case reduces to the index gap. If your pipeline thresholds on the probability — auto-approve above 0.9, escalate below — then calibration is the only number that matters, and it is a number you will have to produce yourself from a labelled sample.

When the doubling pays for itself

Work it through at a volume where the rate difference stops being noise. A moderation gate asking one noul question per post, 500,000 posts a month, at roughly 400 input tokens each, is 200 million input tokens a month.

• On d1 — $4.00 a month. On d1-ultra — $8.40 a month. The four-dollar difference is not a budget line; it is smaller than the cost of the engineer-hour spent deciding.

• Push the same gate to image input and the token counts climb fast, because a single 1024×1024 image is 1,536 input tokens and every question re-bills its text and all of its images. Five questions about the same photo pays for that photo five times, which is a design mistake worth avoiding regardless of which model you pick.

• At 20 billion input tokens a month — a large-scale classifier, not a pilot — d1 costs about $400 and d1-ultra about $840. The absolute gap is now real, and it is still under a thousand dollars for a workload that is processing the equivalent of tens of millions of images.

The conclusion this arithmetic forces is not the one people expect. For the overwhelming majority of decision workloads, the price difference between d1-ultra and d1 is too small to be the deciding factor, which means the decision should be made entirely on the quality of the answers. The interesting exception runs the other way: if your volume is genuinely enormous and your task is easy — a binary gate on short text, no images — you should not be shopping between these two at all, because the open-weight d1-3B does that job at zero per call. The pair's pricing is close enough that the real question is the one below them.

Where a decision model is the wrong instrument

Both models are documented for the same four jobs: classification and routing, scoring and rating, binary gates, and multimodal inspection. The vendor's own documentation is equally clear about what they are not for — free-form generation, multi-turn chat, summarisation, open-ended reasoning, writing code. There are no output tokens to stream, so an application that expects prose back will get an error, not a worse answer. That is worth stating because the temptation to reach for a decision model as a cheap substitute for a generative call is exactly how pipelines end up misusing one.

The published evidence for the older model is at least concrete. Across six real applications tested on October 5, 2026, d1 was reported to match or beat GPT-6.1 Sol on four of them at 19 to 200 times lower cost and to answer faster on every task. One of those six was a visual-inspection run over parts from four production lines — circuit boards, candles, cashews and chewing gum, on the VisA dataset — where the model sorted good from defective at 85 to 97 percent accuracy with no task-specific training. Another removed 52 percent of tokens in a context-compaction pass. Those are vendor-run evaluations with a stated methodology, and they are still vendor-run: one run per model per application, list prices without cache discounts, and several of the code and compaction questions authored after d1's pipeline was already fixed. Read them as evidence that the category works, not as a scoreboard you can transfer to your own task.

The upgrade decision, as a rule

Start on d1. Not because it is the better model — the index claim says it is not — but because at $0.02 per million input tokens with no output charge, a few hundred thousand calls cost pennies, and the fastest way to learn whether a decision model suits your pipeline is to put one in front of real traffic and look at the probability distribution it returns. Log the probabilities, not just the labels. If they cluster tightly around the extremes and your thresholds are stable, you are getting calibration for free and have no reason to pay for more.

Move to d1-ultra when a specific, named failure appears in that log: the hard knowledge question, the visually ambiguous claim, the case where the model returns 0.6 and means it. The migration is a model-string change on the same endpoint, so it can be selective — d1-ultra on the routes that showed the failures, d1 on everything else. That is the honest shape of this comparison. It is not an either/or; it is a per-route setting, and the pricing is close enough that you can afford to be surgical about it.

There is one practical wrinkle worth planning for. If your decisions feed a generative step — a reply, a summary, a tool call — you are maintaining two vendor relationships, not one. OrcaRouter collapses the second half of that: one OpenAI-compatible key across 200+ generative models at provider list price passed through with 0% markup, so a price change upstream lands on your bill the same day rather than at the next contract renewal, with automatic failover across providers behind the same endpoint. To be exact about the boundary: we do not host Liquid AI's decision models, so the d1 or d1-ultra call still goes to Liquid AI's own API, and what we carry is everything downstream of the decision.

What to watch before you commit to d1-ultra at scale

Two numbers, both currently missing. The first is latency: d1's 200-to-300-millisecond figure does not tell you anything about d1-ultra, and the newer model publishes nothing, so any latency-sensitive route should be canaried rather than migrated wholesale. The second is an independent calibration measurement. The 112-evaluation comparison and the eight-point index lead are Liquid AI's own reporting, and the index figure reached this article through the vendor's announcement and press coverage of it rather than from a table we could load ourselves. When someone reproduces a reliability curve for either model on public data, the 2.1-times premium becomes a purchasing decision instead of a judgement call.

Until then, the position that survives scrutiny is a conservative one. d1 is a capable, cheap, proven-by-its-own-vendor decision model at a price that makes experimentation nearly free. d1-ultra is a plausible improvement with a specific, checkable-sounding claim attached to it and an unusually thin spec sheet behind it. Paying double for the second one before you have a failure in hand is buying insurance against a problem you have not yet observed — which, at these volumes, is not a mistake so much as a rounding error. Just do not let the rounding error become the architecture.