Artificial Analysis Decision Index: how Drex Max ranks

The Artificial Analysis Decision Index is an independent benchmark of decision models: models that answer by choosing from a fixed set of options, with a probability for each. Artificial Analysis runs every model itself on six evaluations and combines them into one score. On v2.1, Drex Max v1.1 by Nace.AI scores 72.6, ahead of GLiDE v1.2 at 72.5 and every other decision model.

Where the 72.6 comes from

No decision model wins every evaluation. Drex Max v1.1 ranks #1 of 18 decision models on AA-Paperclip, the most heavily weighted part of the index at 30%, and no lower than #5 on the other 5. That balance puts it first among decision models.

1. AA-Paperclip · 30%

   decisions and extraction over resumes, receipts, invoices and research papers

   72.8#1 of 18

2. EvalSafe-O*NET-AA · 15%

   judgement about workplace documents, from TypeSafe's O*NET set

   88.3#5 of 18

3. AA-Omniscience-Decision · 15%

   knowledge, and declining instead of guessing

   60.3#5 of 18

4. HLE-MultiChoice · 15%

   reasoning on multiple-choice questions from Humanity's Last Exam

   46.6#5 of 18

5. AA-Symmetry · 15%

   whether a decision holds when options are reordered or reworded, or the question is negated

   79.8#4 of 18

6. When2Call · 10%

   whether to call a tool, and which, from NVIDIA's evaluation

   95.3#3 of 18

Best decision-model score on each evaluation. Rank and best score are among the 18 decision models.

$103

per million tasks, against $780 for GLiDE v1.2, the next-highest score

283 ms

median time per task, against 284 ms for GLiDE v1.2

The top decision models

1. 1Drex Max v1.1 · Nace.AI72.6
2. 2GLiDE v1.2 · Fastino72.5
3. 3Mammoth 27B · Base Labs71.4
4. 4d1-ultra · Liquid AI70.6
5. 5Decider v1.2 · Perplexity69.8

Ranked among the 18 decision models on the index. Artificial Analysis can also plot language models alongside them for reference.

What the Decision Index measures

Each answer gets a Decision Score of 50 × (1 + c − (p − c)²), where p is the probability the model gave its chosen answer and c is 1 if that answer is right. A right answer scores between 50 and 100, a wrong one between 50 and 0, so a model loses most when it is confidently wrong. Artificial Analysis describes it as the payoff of acting on answers above a confidence threshold and routing the rest, averaged over every threshold.

30%

AA-Paperclipdecisions and extraction over resumes, receipts, invoices and research papers

15%

EvalSafe-O*NET-AAjudgement about workplace documents, from TypeSafe's O*NET set

15%

AA-Omniscience-Decisionknowledge, and declining instead of guessing

15%

HLE-MultiChoicereasoning on multiple-choice questions from Humanity's Last Exam

15%

AA-Symmetrywhether a decision holds when options are reordered or reworded, or the question is negated

10%

When2Callwhether to call a tool, and which, from NVIDIA's evaluation

AA-Symmetry is scored differently: a model gets nothing if its top option changes when the question is reworded, and otherwise loses the range of its confidence across the versions.

Cost is the average cost per task at current prices, times one million. Decision models are billed for input tokens only; language models for input and output. Time is the median time per task through each provider's API, in milliseconds. Both are averaged across the evaluations with the same weights.

What is a decision model?

A decision model reads a case and a fixed set of options and returns a probability for each option in one pass. A language model writes text that has to be parsed into an answer. Because a decision model returns probabilities instead of tokens, it is fast and cheap to run, it cannot answer outside the options, and how sure it is can be measured. That is what the Decision Index scores.

Drex Max vs Jev 1.14

Jev 1.14 is TypeSafe's decision model. On the Artificial Analysis Decision Index v2.1, as of Nov 12, 2026, Drex Max v1.1 scores 72.6 against 66.2. Jev 1.14 is ahead on EvalSafe-O*NET-AA and answers faster, at 92 ms per task against 283 ms.

Higher is better on every row; the higher score is in bold.

The Public Decision Index (0.4.1)

The Public Decision Index is a different leaderboard: a community board on Hugging Face, named after Jev, that scores chance-corrected decisions across public benchmark suites. It is versioned by edition rather than by date. On edition 0.4.1, where Drex Max is listed as Drex Max v1, its public score is 67.5 against 60.7 for Jev 1.14.

Frequently asked questions