A generated comparison card headed 'Drex Max v1.1 vs Liquid AI d1-ultra', with rows for Decision Index 73.4 against 70.4, document decision score 72.8 against 64.7, abstention rate 65.3 percent against 88.6 percent, time per task 275 ms against 295 ms, and cost per 1M tasks $102 against $132, with a footer reading Artificial Analysis Decision Index v2, 8 November 2026.
Guides & Insights

Drex Max v1.1 vs Liquid AI d1-ultra: Two Multimodal Scorers, and a Headline That No Longer Holds

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Liquid AI d1-ultra shipped on 15 October 2026 with a sentence that was defensible at the time: it "outperforms all other decision models on the Artificial Analysis Decision Index by at least 8 points." Three weeks later Artificial Analysis replaced that index with v2, re-scored seventeen models on new data, and d1-ultra landed fifth at 70.4 — behind Drex Max v1.1 at 73.4, GLiDE v1.2 at 72.8, Mammoth 27B at 72.4 and, by three hundredths of a point, Perplexity's Decider v1.2 at 70.5. The eight-point lead is gone, and it was never measured on this instrument.

That is the frame for this comparison, and it is worth being precise about what it does and does not mean. It does not mean d1-ultra got worse. It means the board changed underneath a claim that was made against the old one, which is the ordinary hazard of quoting a leaderboard position as though it were a property of the model. Both of these are strong multimodal decision scorers. They are also closer than either vendor's marketing implies, and they are good at visibly different things.

Where the two models agree, which is most of the surface

Both are decision models in the strict sense: you send a state and typed questions, one forward pass returns a probability per option, and no text is generated. Both take images alongside text. Both answer the same three question shapes — a choice among named options, a yes/no, and a score on a scale. Both bill input only, with output unbilled, because there is no output.

• Modality — Drex Max v1.1 reads text, screenshots, charts, forms, scanned documents, photos and video frames, eight images per request. d1-ultra reads text, images, or both; Liquid's launch example is a photograph of a circuit board checked against its spec.

• Price per million input tokens — $0.03 for Drex Max v1.1 during its launch period (50% off a $0.06 list), $0.042 for d1-ultra. Nace is cheaper today and may not be after the discount closes.

• Image billing — Drex Max v1.1 counts image tokens inside the 131,072-token per-question limit. d1-ultra bills 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image costs 1,536 tokens, and bills each question as its own prompt including all images.

• Context — 131,072 tokens per question for Drex Max v1.1, with 262,144 for the whole packed request. Liquid does not publish a comparable per-question window.

• Availability — Drex Max v1.1 sells through Nace's API and console; d1-ultra through the Liquid API and console. Neither is on our route list, and nothing here is an availability claim from us.

That per-question billing rule on d1-ultra is the detail that decides cost at document scale, and the board measured it. Artificial Analysis reports Drex Max v1.1 at about 9,000 input tokens for a fifteen-question document task against roughly 63,000 for models that re-bill the input per question — and d1-ultra's per-task cost came out at $132 against Drex's $102, close to the ratio of their list prices and slightly worse.

The board, component by component

The composite gap is 2.9 points. The component gaps are wildly uneven, and the pattern is consistent: Drex wins on documents, d1-ultra wins on knowing things and on declining to answer.

Screenshot of the Artificial Analysis Decision Index v2 page, showing the leaderboard navigation and the board's stated definition - a composite index testing decision models across document decisions and extraction, tool calling, knowledge, reasoning and invariance, built from six evaluations including AA-Paperclip's 1,500 resumes, receipts and invoices.

• Decision Index v2 — Drex Max v1.1 73.39, d1-ultra 70.45.

• Accuracy index — Drex Max v1.1 58.41, d1-ultra 54.94, which is the score with the calibration component stripped out.

• AA-Paperclip document decision score — Drex Max v1.1 72.82, d1-ultra 64.69. All-fifteen-correct rate: 63.0% against 51.3%.

• When2Call tool calling — Drex Max v1.1 95.26, d1-ultra 93.83. Close, and both far above Jev 1.14's 84.39.

• AA-Symmetry paraphrase consistency — Drex Max v1.1 84.96, d1-ultra 77.81, a 7-point gap that is the second-largest divergence in the set.

• AA-Omniscience-Decision — d1-ultra 64.39, Drex Max v1.1 60.27. The one evaluation where the ordering flips.

• HLE-MultiChoice — d1-ultra 47.39, Drex Max v1.1 46.61, but on raw accuracy d1-ultra is at 6.09% and Drex at 8.25%.

• Abstention — d1-ultra declines on 88.61% of omniscience questions it cannot answer, Drex Max v1.1 on 65.28%, Microsoft-Decision-1 on 21.74%.

• Time per task — Drex Max v1.1 275 ms, d1-ultra 295 ms.

• Cost per 1M tasks — Drex Max v1.1 $102, d1-ultra $132.

Read the abstention row twice, because it explains the omniscience row and it is the least intuitive property on the board. d1-ultra's higher omniscience decision score is not mainly knowledge: its raw accuracy there is 28.0%, the same as Drex Max v1.1's. It scores higher because it says "I don't know" far more often and the scoring function pays for that. Mammoth 27B and Elephant 9B lead the symmetry evaluation for the same reason — Artificial Analysis says outright that Baseten's models "consistently admit they do not know an answer with high confidence across question variations."

None of that is a flaw in the metric. It is a statement about what you are buying. d1-ultra is the model to reach for when an unanswered question is cheaper than a wrong one. Drex Max v1.1 is the model to reach for when something has to come back with a decision, and the low-confidence tail gets a human.

A generated scoreboard headed 'Two multimodal scorers, five uneven gaps', listing Drex Max v1.1 against Liquid AI d1-ultra on document decision score at 72.8 versus 64.7, paraphrase symmetry at 85.0 versus 77.8, tool calling at 95.3 versus 93.8, knowledge and abstention at 60.3 versus 64.4, and abstention on unanswerable items at 65.3 percent versus 88.6 percent.

What Liquid's own evaluation measured, and what it did not

Liquid's launch post compares d1-ultra to d1 and Jev 1.14 across "112 unique evaluations" and reports that it beats Jev on every one, with better calibration at the same price. That is a vendor-run comparison on a vendor-chosen instrument, and it says nothing about Drex Max, which does not appear in it. The eight-point index claim was made against Decision Index v1.0, published 13 October 2026 — two days before d1-ultra existed — and that board never listed d1-ultra.

The vendor does make one claim the new board corroborates: it trained d1-ultra specifically to be less overconfident, and the calibration showing up as a 64.39 on omniscience against Jev's 60.14 and a 2.0-point higher composite than Jev at comparable accuracy is real. d1-ultra at 70.45 against Jev 1.14 at 66.96 on the same instrument is a genuine improvement over the incumbent, measured independently.

Screenshot of Liquid AI's d1-ultra announcement dated October 15, 2026, headed 'Introducing the World's Best Decision Model', describing the model as built for harder real-world cases such as spotting a defect on a circuit board against its spec or checking a damage claim against a customer photo, and stating availability on the Liquid API at $0.042 per million input tokens.

The claim that does not survive is the ranking. On v2, d1-ultra is fifth of seventeen, and the four models above it are all inside 3 points. A 3-point spread across the top five is not a hierarchy; it is a tie broken by which component your workload happens to weight.

What to actually run

Both models are hosted, neither is on our catalogue, and the honest way to choose between them is a bake-off on your own states rather than on either vendor's suite. If you run it, run it on the thing the board penalises: all-or-nothing documents. A model that gets 97% of individual answers right and 63% of documents fully right is not twice as good as one at 51%; it is exactly as good as the probability that none of fifteen dependent answers is wrong, which is what your pipeline will experience.

What a router changes here is not the choice between these two — we carry neither — but the shape of the pipeline around them. Both are scorers that sit upstream of something that writes. The writing step is a general model call, and that is the part where one endpoint with 200+ models, provider list price passed through at no per-token markup and automatic failover actually saves you an integration, because the model you route to will change more often than the scorer will.

If you want a single reason to prefer Drex Max v1.1 today, it is not the 2.9-point composite. It is that it is faster, cheaper per task, and seven points more consistent when the same question is worded two ways — and consistency is the property that breaks quietly in production rather than loudly in a benchmark. If you want a single reason to prefer d1-ultra, it is that it declines. Both are defensible. Neither vendor's headline is.