Liquid AI d1-ultra: The New Decision Index Leader, Trained to Doubt Itself

Liquid AI d1-ultra shipped on October 15, 2026, and the most interesting thing about it is a number nobody is quoting: on the Artificial Analysis Decision Index, the model it replaces scored 2 on non-hallucination. Liquid AI d1 was, on the same board, the single most accurate decision model at answering knowledge questions — and the worst in its class at recognising the questions it could not answer. Liquid AI d1-ultra is the vendor's answer to that, and it is being announced with a first-place claim: that it tops the Decision Index by at least eight points over every other decision model, including the one currently on top of that board, Jev 1.13. That claim is Liquid's, not Artificial Analysis's, and the board it refers to was published before this model existed. Both facts are load-bearing, and this piece keeps them attached to the numbers they belong to. Every provider outage and rollback we have tracked is logged in the Orca AI Incident Archive, which carries the full record set alongside this incident's neighbouring entries.

What the model is, mechanically, is the same bet Liquid has been making all year. You give it a state — a JSON payload, a support ticket, a dashboard screenshot — and a set of named questions, and it reads answers off its own distribution in a single forward pass. No sampling, no parsing, no runaway generation. The vendor's usage counter reports output_tokens: 0, exactly as it does for the rest of the d1 line.

What Liquid says it built, and what it tested it against

Every number in this section is Liquid's own, reported in the launch post, unreproduced by anyone else as of this writing. The model is described as a multimodal decision model built on a larger open-weight base with what the post calls heavy Liquid-specific customisation, tuned specifically toward knowledge-heavy decisions — a phrase that reads as a direct reference to the calibration problem below. It carries the same three question types as the rest of the family. noul is binary: give it a statement and it returns P(yes) between 0 and 1. Choice picks a label and returns a distribution plus a confidence. Score places the state on an ordered scale with per-level probabilities. All three come back structured, and none of them cost output tokens.

Liquid's evidence is a run of 112 unique evaluations against its own earlier d1 and against Jev 1.14, the newer Jev release. The company says d1-ultra outperforms Jev 1.14 across all tasks and is better calibrated at the same price. That is a stronger and more specific claim than "leads the leaderboard," and it is also the harder one to check, because calibration is the property you cannot read off a single score. Note also that Jev 1.14 is a vendor-named comparator: it appears nowhere on our catalogue, and we could not locate a public Decision Index row for it.

The API surface is a single non-streaming call to POST https://api.liquid.ai/decisions/v1/systemone with "model": "d1-ultra". There is no context window published anywhere in the launch material, which is a real gap rather than a rounding error: for a model whose whole job is to be handed state, the size of the state you are allowed to hand it is close to the primary spec. Since it never generates tokens, output cost is structurally zero and streaming does not apply.

The index in question predates the model

Artificial Analysis published the Decision Index v1.0 on October 13, 2026 — two days before d1-ultra existed. The board is built from four evaluations weighted at 25% each: AA-Omniscience-Decision, a private set of 500 questions scored half on accuracy and half on non-hallucination; HLE-MultiChoice, 513 text multiple-choice questions drawn from Humanity's Last Exam; When2Call, an N​VIDIA tool-calling evaluation; and WildGuardTest, Ai2's safety classification set.

The article that introduced it does not mention d1-ultra anywhere, which is unsurprising, and it is why the "eight points clear" figure should be read as a vendor claim rather than a ranking. On the published board, Jev 1.13 leads the decision-model field at 58, with GLiDE and M​icrosoft-Decision-1 tied at 54. Two language models are placed above all of them: C​laude H​aiku 5.5 (Low) at 61 and GPT-6 Luna (Low) at 55. Run the arithmetic on Liquid's claim and it lands at 66 or higher — above the best language model on the board. That may well be what the model does. It is also exactly the kind of claim that needs a third party to reproduce it before anyone plans around it.

Why calibration, specifically, is the whole story

The failure d1-ultra is aimed at is the one Artificial Analysis named in the same article. Decision models, as a class, were found to be broadly overconfident: the probability a model states for an answer routinely runs ahead of how often that answer turns out to be right. That matters more here than it would for a chatbot, because the entire output of a decision model is a probability. A confidently wrong language model produces a paragraph you can scrutinise. A confidently wrong decision model produces a number a pipeline is designed to act on.

The spread inside the class is wide, and d1 sat at the bad end of it:

• AA-Omniscience-Decision accuracy — Liquid AI d1 was the top decision model at 41, with Jev 1.13 at 35. Language models on the same evaluation scored 52 and 66.

• Non-hallucination rate — Liquid AI d1 scored 2. Jev 1.13 scored 35, Mercury Decide 32, and M​icrosoft-Decision-1 led at 40. No other decision model on the board failed to reach 17.

• The gap between those two rows is the whole pitch. d1 knew the most and admitted the least. A model trained to close that gap — the one dimension where Liquid's previous entry was last rather than first — would move the class forward even without touching raw accuracy.

• HLE-MultiChoice — Jev 1.13 led at 21, every other decision model at 15 or below. This is the eval where the field has the most headroom, and the one Liquid's 112-evaluation run does not report against.

• When2Call and WildGuardTest — Celeris-1 Decision and GLiDE tied at 84 on tool-calling, with Jev 1.13 fourth at 81; GLiDE led safety at 95, with Decider v1.1 27B and Jev 1.13 together at 94.

What a million decisions actually costs

Pricing is the part of the launch that needs no caveat. Liquid AI d1-ultra is billed at $0.042 per million input tokens and nothing at all for output, because there is no output. That price is identical to d1's and identical to what we charge for Jev 1.13, which makes the two models directly comparable on cost and nothing else.

The arithmetic is simple enough to do on a napkin. A state of roughly a thousand tokens — a moderately detailed JSON payload, or a page of text — costs $0.042 to run one million times. Artificial Analysis's own cost-per-million-tasks figures for this class land between $17 and $199, with Jev 1.13 at $42, M​icrosoft-Decision-1 at $31 and Decider v1.1 27B at $17; the per-task median time across the class runs from 0.06 seconds to half a second. A thousand-token decision at $42 per million tasks sits in the middle of that band, which is the useful thing to know: the pricing is unremarkable for the category, and the interesting question is not cost per token but cost per correct decision.

There is one billing detail that will catch people, and it is worth reading twice. Images are charged as input text at 1.5 tokens per 32×32-pixel patch, making a 1024×1024 image 1,536 tokens — and every question is billed as its own prompt, including all attached images. Ask ten separate questions about the same photograph and the image is billed ten times, so those ten questions carry 15,360 image tokens before a single word of text. Batching your questions into one call does not amortise it. If your workload is visual, the number that governs your bill is questions multiplied by image tokens, not documents multiplied by image tokens.

Where you can call it

Liquid AI d1-ultra is live now on Liquid's own API, through the Liquid console, and in the d1 Playground for anyone who wants to poke at it without writing a client. The vendor has said additional platforms are coming; the honest summary today is the vendor's own API, with several third-party platforms to follow. There is a code sample in the launch post that will run as-is.

A routing note, and it is not a pitch: Liquid AI d1-ultra is not on our catalogue, and this article is not an availability claim for it. What is on our catalogue is Jev 1.13 — the model that currently leads the Decision Index — at the same $0.042 per million input tokens and the same zero completion cost, with a 64K context window, served over the identical systemone endpoint shape: text in, structured JSON out, non-streaming. That is a useful accident for anyone evaluating this launch, because it means the challenger and the incumbent can be run side by side without a second contract. d1-ultra goes to Liquid directly; the incumbent is already on our catalogue. Both sit behind one API for 200+ models, where list prices are passed through at 0% markup and where automatic failover means a new model can take a slice of a real workload without being trusted with all of it.

That last part is the practical recommendation rather than a product placement. A model that is one day old, whose headline benchmark has not been independently reproduced, and whose context window is unpublished is not a thing to route production traffic to on day one. It is a thing to shadow-run against the model already in the slot. Run the same states through both, compare the distributions rather than the labels, and watch specifically for the case where d1-ultra says 0.8 and turns out to be wrong — because that is the case the launch is entirely about, and it is the one number Liquid has not published.

What would settle it

Three things, in descending order of how much they would change the decision.

The first is a Decision Index row. Until Artificial Analysis runs d1-ultra through its own pipeline, the eight-point margin over Jev 1.13 is a vendor figure, and the calibration improvement over Jev 1.14 is a vendor figure with no public counterpart at all. The board is updated regularly, and a decision model that genuinely clears 66 would be the most interesting result in the category this year, because it would put a non-generative model above every language model on the same index.

The second is the missing spec sheet. Context window, latency at a stated batch size, and whether questions can share an image without re-billing it — those three numbers decide whether a workload is feasible long before benchmark scores do. Liquid's predecessor generation ran text decisions in 200 to 300 milliseconds, which is a useful reference point but not a measurement of this model.

The third is calibration outside the lab. Liquid's 112 evaluations are the vendor's own selection of tasks. Overconfidence is a property that shows up in the tail — the rare hard question, the ambiguous state, the input from a distribution the model never saw. That is where d1 scored 2, and it is the only place where a real improvement would be visible.

Until then the honest position is the boring one: the price is real and public, the endpoint works, the category's worst flaw is correctly identified, and the fix is claimed rather than demonstrated. If you work on structured decisions, it is worth an afternoon. It is not worth a migration.