A generated title card for Drex Max v1.1 subtitled 'first on the Artificial Analysis Decision Index v2, by 0.56 points', with chips reading 73.4 Decision Index, 275 ms per task, $102 per 1M tasks and a 262,144-token request limit, plus a leaderboard strip showing 73.4, 72.8 and 72.4 for the top three decision models.
Guides & Insights

Drex Max v1.1 Tops the Decision Index — and the Board's Own Details Qualify the Win

Author

Gideon Frost

Date Published

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

Drex Max v1.1, the multimodal decision model from Nace.AI, sits first on the Artificial Analysis Decision Index v2 with a score of 73.4. The margin is 0.56 points over GLiDE v1.2 from Fastino Labs at 72.8, and 0.94 over Mammoth 27B at 72.4. Those three numbers are the whole story of the board's top, and they are also the reason the headline needs a second sentence: the same evaluation says GLiDE is more accurate than Drex Max v1.1, and that Mammoth 27B holds the same answer across paraphrases far more often.

The model itself is not new today. Nace's own model reference lists drex-max-v1.1 as released on 28 October 2026, and the retired drex-max id has been answered by v1.1 since 8 November. What is new is the measurement. Artificial Analysis published Decision Index v2 on 8 November 2026, two days before this piece, and it is the first independent board that scores this checkpoint against the rest of the decision-model field on one instrument. That is the news: not a launch, but the first outside number for a model that had only shipped its vendor's own.

What the board actually measures now

Decision Index v2 is a different instrument from the one that produced the 64 that Drex Max carried in October, so the two are not comparable and no one should line them up. The v2 index is a weighted average of six evaluations, and 45% of it is now built on private data, up from 25% in v1 — a deliberate move against benchmark overfitting.

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

• AA-Paperclip, 30% of the index — 1,500 private documents (resumes, receipts, invoices, research papers), each carrying five decisions and ten facts to extract, scored as one document that only counts if all fifteen answers are right.

• EvalSafe-O*NET-AA, 15% — workplace-document judgement, built on T​ypeSafe's O*NET benchmark.

• AA-Omniscience-Decision v1.1, 15% — knowledge and hallucination, with the "I don't know" option reworded ten ways.

• HLE-MultiChoice v1.1, 15% — Humanity's Last Exam multiple choice, also with a decline option.

• AA-Symmetry, 15% — 120 questions expanded to 440 variations by reordering options, swapping labels and negating yes/no questions.

• When2Call, 10% — NVIDIA's tool-calling evaluation, down from 25% in v1.

The headline metric is new too. Each question scores 50 × (1 + c − (p − c)²), where p is the probability the model put on its answer and c is whether that answer was right. In aggregate that is 50 × (1 + accuracy − Brier score). A model is paid for being right and for knowing how right it is, and it is punished more for a confident error than a hesitant one. That matters here, because Drex's pitch is calibrated probability rather than prose, and this is the first board that scores the calibration directly instead of treating it as a footnote.

Where Drex Max v1.1 wins, by how much, and against whom

The composite hides a split field. Pulled apart, Drex Max v1.1's lead comes from two components out of six, and one of them is the one Nace built the model for.

• AA-Paperclip (documents) — Drex Max v1.1 scores 72.8 on the decision score with 63.0% of documents fully correct; d1-ultra from Liquid AI is at 64.7 and 51.3%, Microsoft-Decision-1 at 51.8 and 20.9%.

• When2Call (tool calling) — Drex Max v1.1 95.26 against Jev 1.14's 84.39 and Microsoft-Decision-1's 87.95, the widest single-evaluation gap in the comparison set.

• AA-Symmetry — Drex Max v1.1 84.96, Jev 1.14 79.31, d1-ultra 77.81, Microsoft-Decision-1 75.25. GLiDE v1.2 is at 84.18, effectively level.

• AA-Omniscience-Decision — Drex Max v1.1 60.27, d1-ultra 64.39. This one goes the other way, and it is the evaluation about knowing things rather than deciding things.

• HLE-MultiChoice — Drex Max v1.1 46.61 against Jev 1.14's 44.47, but d1-ultra is at 47.39 and the raw accuracy is 8.25% against Jev's 12.77%.

• Time per task — 275 ms for Drex Max v1.1, 92 ms for Jev 1.14, 295 ms for d1-ultra. Drex is the fastest frontier-scale decision model on the board and roughly three times slower than the cheapest-to-run incumbent.

A generated component breakdown headed 'Drex Max v1.1 vs GLiDE v1.2 vs Mammoth 27B — where the points come from', with five weighted rows: document decisions at 30 percent weight scoring 72.8, 76.0 and 66.0; workplace judgement at 15 percent scoring 88.3, 84.1 and 79.9; tool calling at 10 percent scoring 95.3, 90.1 and 87.0; paraphrase symmetry at 15 percent scoring 85.0, 84.2 and 99.0; and knowledge and abstention at 15 percent scoring 60.3, 58.6 and 59.0, footnoted to the Artificial Analysis Decision Index v2 of 8 November 2026.

Four qualifications the board publishes and the press release does not

Artificial Analysis is unusually explicit about what its own top score does not mean, and the article announcing v2 says each of these in plain language.

The first is accuracy. "GLiDE v1.2 leads Drex Max v1.1 on accuracy metrics (60 vs 58), while Drex Max v1.1 edges ahead on the Decision Index." Drex wins the composite because it is better calibrated, not because it gets more answers right. If your pipeline throws away the probability and keeps only the argmax, you have bought the part of the model that the board did not reward.

The second is consistency under paraphrase. Mammoth 27B selects the same answer across AA-Symmetry's variations on 99% of questions; Drex Max v1.1 does so on 91%, GLiDE on 88%. Drex is still well ahead of where it was, and ahead of Jev, but "the same question worded two ways gets the same answer" is a property Mammoth has and Drex does not fully have.

The third is the all-or-nothing document test. Drex Max v1.1 gets 97% of individual answers right on AA-Paperclip but passes only 63% of documents, because one wrong answer out of fifteen fails the document. It makes all five decisions correctly on 71% of documents and all ten extractions on 88%. Claude Haiku 5.5 (Max), a general-purpose language model, passes 92%. The gap between 97% of answers and 63% of documents is the cost of chaining fifteen dependent calls without a checker.

The fourth is abstention. On the omniscience evaluation, d1-ultra declines rather than guesses on 88.6% of the questions it cannot answer; Drex Max v1.1 on 65.3%; Microsoft-Decision-1 on 21.7%. Drex is a model that would rather commit than abstain, which is fine when a human reviews the low-confidence tail and expensive when nothing does.

The cost column is the strongest thing Nace has

Cost per 1 million tasks on the same board: Celeris-1.1 Decision $14, Decider v1.2 $68, Drex Max v1.1 $102, d1-ultra $132, Microsoft-Decision-1 $142, Jev 1.14 $156, GLiDE v1.2 $780, Clef $859. Above the decision models, GPT-6 Luna (Max) costs $1,530 and Claude Haiku 5.5 (Max) $6,240 for three to four index points.

Part of that spread is token price and part is billing shape, and the board separates them. Drex Max v1.1 and Decider v1.2 evaluate a document once and answer every question from that single pass, so an AA-Paperclip task uses roughly 9,000 input tokens. GLiDE v1.2 and Decider Flash v1.2 bill the input again for each question, about 63,000 tokens per task, which is how GLiDE lands at $780 on a $0.04 list price — nearly eight times Drex's cost at a 33% higher token rate. Nace's own table reports "$0.03 per 1,000 decisions" and the board reports $102 per 1 million tasks; those are different units measuring different workloads, and only the second one lets you compare across models.

Nace's product page also headlines "118× faster than Claude Haiku 5.5 (Max)". Treat that as a vendor claim on a vendor denominator. The board's unit is time per task, where Drex Max v1.1 is 275 ms and AA's write-up puts GPT-6 Luna (Max) at 11.34 s — roughly 40×. Both numbers can be true at once because one counts a single decision and the other counts a whole fifteen-question document. Plan capacity from the per-task figure.

The id is the part that will break your integration

Three things in Nace's documentation will cost an afternoon if you find them after wiring the client.

• drex-latest does not point at a Max model. It resolves to drex-v2.0, an open-weights text-only mixture of experts that scores 63.43 on the same board. A request that omits model entirely is also answered by drex-v2.0. If you copied a quickstart that sets TYPESAFE_DEFAULT_MODEL=drex-latest — and Nace's own Jev-compatibility instructions tell you to — you are calling the wrong model and will see a 10-point index drop with no error.

• drex-max is retired as of 8 November 2026 and is served by drex-max-v1.1 at v1.1's limits and price. Existing clients keep working; the response reports the successor's id. New work should send drex-max-v1.1 explicitly.

• The context figure is stated twice, differently. The API guide says 131,072 tokens per question with image tokens counted; the model reference says a 262,144-token input limit for the whole packed request, including state, all eight images, every question and the model's formatting. Size for the smaller number unless you have tested the larger.

Screenshot of Nace's Drex model reference page, showing the documentation sidebar covering the Drex API, guides, the Drex Max section and the Drex reference pages for Models, Errors, Limits and Pricing, with the Models page open and described as the table of models Drex serves, their API ids, prices and token limits.

None of these is a defect. They are the ordinary debris of in-place versioning, and they are the reason the release date on the model matters more than the version number in the name.

What this means if you are choosing today

Drex Max v1.1 is on no route we serve. The OrcaRouter catalogue does not carry it, and nothing here should be read as an availability claim — Nace sells it through its own API and its own console, at $0.03 per million input tokens during the launch period, 50% off a $0.06 list. What the catalogue does carry is the incumbent this model is measured against: TypeSafe Jev 1.14 is a listed route at $0.042 per million input tokens, with the same typed-question request shape and a 66.96 on the same board. Running both means two contracts and two SDKs, which is the real cost of a bake-off, and it is the usual reason people put the comparison behind a router instead.

The honest reading of 73.4 is narrower than "best decision model". Drex Max v1.1 is the best composite score on a board that is 45% private and two days old, it is the fastest model at that score by a wide margin, and it is third-cheapest on the Pareto frontier. It is not the most accurate, it is not the most consistent, and it is not the most willing to abstain. Which of those four matters depends entirely on whether a human sees the low-confidence tail — and that question is not on any leaderboard.

What would change this read: a second independent board reproducing the 73.4, or a v1.2 that lifts HLE-MultiChoice accuracy above Jev's 12.77%. Neither exists yet. Until one does, the number to quote is 73.4 with the word "narrow" attached.