A generated title card headed ‘2.2 points apart, and a price that doubles’, subtitled ‘d1.2-ultra vs Drex Max v1.1 - Decision Index v2.1’. Two rounded cards read ‘d1.2-ultra’ over ‘75’ and ‘Drex Max v1.1’ over ‘73’, separated by a thin vertical divider; below them a wide strip reads ‘$133 per 1M tasks vs $103 now, about $206 at standard price’, and one outlined chip reads ‘independent board, both models run by Artificial Analysis’. The OrcaRouter mark is composited bottom right.
Guides & Insights

d1.2-ultra vs Drex Max 1.1: A 2.2-Point Gap and a Price Tag That Doubles in December

Author

jinhao.song

Date Published

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

Liquid AI d1.2-ultra leads Nace.AI's Drex Max v1.1 by 2.23 points on Artificial Analysis Decision Index v2.1 — 74.85 against 72.62 — and on sticker price it is the more expensive model, at $0.042 per million input tokens against $0.03. Both of those numbers are true today and only one of them will still be true when the launch discount on the Nace side expires. Artificial Analysis attaches an asterisk to Drex Max v1.1's cost row: the $103 per million tasks it reports "reflects a temporary launch price," and at the standard price that cost would double. Which means the pair that looks like a premium model against a value model is, on the cost that actually governs a deployment, the other way round.

The ranking is real and independently measured — both models were run by Artificial Analysis on its own hardware, not by either vendor — so this is a comparison of two scored rows rather than two press releases. What follows separates the points that will survive the price change from the ones that will not.

First, what the footnote does to the cost comparison

A generated two-column scoreboard titled ‘d1.2-ultra vs Drex Max v1.1 - the scoreboard’. The d1.2-ultra column reads Decision Index 75, AA-Paperclip 75.8 at 67.7% all-pass, AA-Symmetry 86.3, HLE accuracy 2.36%, context window 66k and cost per 1M tasks $133. The Drex Max v1.1 column reads Decision Index 73, AA-Paperclip 72.8 at 63.0% all-pass, AA-Symmetry 79.8, HLE accuracy 8.25%, context window 262k and cost per 1M tasks $103 at launch price. The footer reads ‘Both figures per Artificial Analysis Decision Index v2.1; Drex cost reflects a temporary launch price and would double at standard pricing.’

Per-token prices are the wrong number to plan on when one of the two vendors is discounting. Per-task cost is the right one, because it folds in how many tokens each model burns to answer the same document, and that is what Artificial Analysis publishes alongside the index.

• Per million input tokens — d1.2-ultra $0.042, Drex Max v1.1 $0.03.

• Per million tasks at current prices — d1.2-ultra $133, Drex Max v1.1 $103, the latter flagged as a temporary launch price.

• Per million tasks at standard price — d1.2-ultra $133, Drex Max v1.1 roughly $206 by the board's own note that its cost would be double.

• Class median — $135 per million tasks, so the Liquid model sits just under it and the Nace model will sit well above it.

That inversion matters more than the two-cent token gap suggests, because it changes which model you reach for at volume. The Drex line has been the cheap option in this category since v2 launched on 8 November; if the discount is a launch promotion rather than a repricing, the cheap option becomes the expensive one without a single benchmark moving. Anyone who pinned a budget to the $103 figure should ask Nace when the standard price takes effect before they commit a quarter's traffic to it.

What the 2.2 points are actually made of

A headless-browser screenshot of the Artificial Analysis model page for Drex Max v1.1, captured 17 November 2026 in English, showing the Nace.AI vendor label, the heading ‘Drex Max v1.1: Release, Decision Index Score, Performance & Price’, a Decision Index score of 73, a 262k-token context window, and a summary describing the model as amongst the leading decision models and reasonably priced, but notably slow.

The composite is close. The components are not, and they point in different directions, which is the useful part.

• AA-Symmetry, 15% — the widest gap of the six. d1.2-ultra scores 86.34 against Drex Max v1.1's 79.82. On the stricter all-pass measure, which requires the same answer across paraphrases rather than a good average, it is 56.7% against 53.3%. Drex Max v1.1 was already the second-best model in its own family on this test; the Liquid model is simply more stable under rewording.

• AA-Paperclip, 30% — 75.81 against 72.82 on the decision score, and 67.7% against 63.0% of documents fully correct. Three points of document all-pass rate is the single largest contributor to the index gap, and it is the one that maps most directly onto a real extraction pipeline, because a document that misses one of fifteen answers counts as a failure either way.

• When2Call, 10% — 96.49 against 95.26. Both are near the ceiling of the evaluation; a point and a quarter on a test where the whole top ten lives inside a two-point band is not a differentiator.

• EvalSafe-O*NET-AA, 15% — 89.53 against 88.32, likewise a coin-flip margin.

• AA-Omniscience-Decision v1.1, 15% — 60.80 against 60.27. Effectively tied on the score, and split underneath it: Drex answers more of these questions correctly, 28.0% against 23.6%, while d1.2-ultra abstains more honestly, 83.8% non-hallucination against 65.3%.

• HLE-MultiChoice v1.1, 15% — 46.36 against 46.61, the one decision score Nace wins, by a quarter of a point. Underneath it the raw accuracy gap is not a quarter of a point: 8.25% against 2.36%.

Where Drex Max 1.1 still wins, and why it is not a rounding error

Three things on the board favour the Nace model, and the first two are the kind that decide whether a workload is possible at all.

The context window is the big one. Drex Max v1.1 accepts 262k tokens; d1.2-ultra accepts 66k, which is itself double what Liquid's previous flagship carried. That is a four-fold difference in the size of the state you can hand the model in one call, and neither model generates output tokens, so there is no output budget to trade against it. A 200-page contract, a long claims file, or a multi-document case bundle that fits in one Drex request has to be chunked, re-questioned and reassembled on the Liquid side — and every chunk boundary is a place where a cross-document question stops being answerable.

The second is knowledge. On HLE-MultiChoice, Drex answers 8.25% of questions correctly against 2.36%, and on AA-Omniscience-Decision it answers 28.0% against 23.6% — while abstaining far less. d1.2-ultra's high non-hallucination rates are partly a virtue and partly a symptom: a model that says "I don't know" three times in four is well calibrated and not very useful on a question that genuinely has an answer. If your task is decisions about documents, that trade is fine. If it is decisions that require knowing something, Drex is the model that knows more.

The third is speed, on paper only. 283 ms against 218 ms per task on the index, both slower than the 209 ms class median, and both fast enough that neither will be the bottleneck in a pipeline that does anything else.

The two models are not substitutable, which simplifies the choice

No amount of index arithmetic makes these interchangeable, and pretending otherwise is where most comparison pieces go wrong. One is a 66k-token document reader tuned for calibration and paraphrase stability; the other is a 262k-token reader that discounts its price and answers more hard questions. The crossover is not a score, it is a state size.

• If your states fit in 66k and you threshold on the probability — auto-approve above 0.9, escalate below — d1.2-ultra is the better instrument. Best Brier score of the 24 models on the board at 0.150, against 0.160 for Drex, and a symmetry score six and a half points higher.

• If your states run past 66k, or your questions require knowledge rather than reading, Drex Max v1.1 is the only one of the two that will do the job at all, and the index gap stops being the deciding number.

• If you are cost-sensitive at volume and the Nace price holds, Drex is cheaper per task today. If it reverts, it is the more expensive of the two and the Liquid model wins on both score and cost.

Testing both without betting on either

Neither model is on our catalogue, and nothing here is an availability claim for either: d1.2-ultra is served from Liquid's own API, and Drex Max v1.1 from Nace's. That is exactly the situation where a routing layer earns its keep rather than being a line item, because the question a team actually has — does the 2.2-point gap reproduce on my documents? — is answered by running the same states through both, not by reading a leaderboard.

The third model worth putting in that test is Jev 1.14, which we do host at $0.042 per million input tokens over the same systemone request shape. It leads EvalSafe-O*NET-AA at 91.14, the one component where neither of these two is first, and it answers in 92 ms — the fastest of the three by a wide margin. A three-way shadow run across one key, with the routing DSL sending the same payload to each and automatic failover keeping a production path alive if one vendor has a bad afternoon, is a cheaper way to settle this than a procurement cycle.

What to watch, and when

Two dates decide this matchup. The first is whenever Nace's launch price ends: at $103 per million tasks Drex Max v1.1 is the value pick, at roughly $206 it is the premium one, and the index score does not change either way. Ask before you budget.

The second is the next index revision. Decision Index v2.1 is recent enough that the board still carries a note that it was updated on 16 November, and a 2.2-point gap between first and second is inside the range that one component revision can move. If AA-Paperclip's private document set is refreshed, the 30%-weighted component that supplies most of the gap is the one that will move first.

Until then the honest summary is this: d1.2-ultra is the better model on the evidence, by a margin that is small on the composite and large on symmetry and document all-pass rate, and it is the cheaper model per task the moment the Nace discount expires. Drex Max v1.1 remains the only one of the two that can swallow a 200k-token state, and on questions that need knowledge rather than reading it is still the stronger of the pair. Pick by state size first, price second, and index third.

A headless-browser screenshot of the OrcaRouter model page for typesafe/jev-1.14, captured 17 November 2026 in English, showing the /v1/systemone endpoint, Zero Data Retention availability, an input and output list price of $0.04, and measured latencies of 182 ms and 372 ms.