Artificial Analysis Decision Index v2: Drex Max v1.1 leads 17 decision models at 73, Jev 1.14 scores 67
Artificial Analysis has moved its decision-model leaderboard to Decision Index v2. The methodology page dates v2.0 to November 2026 and v1.0 to October–November 2026. The first public mention we found is an Examenos post on 8 November. The scores below were read from the live leaderboard on 8 November 2026 at 08:50 BRT.
This board is separate from the community Decision Index on Hugging Face and from v1.0 of the same index (our launch note). Scores on the three are not comparable.
What changed from v1.0
v2 has six evaluations instead of four:
| Evaluation | What it tests | Weight |
|---|---|---|
| AA-Paperclip (new, private) | 1,500 resumes, receipts/invoices and research papers; 15 answers per document, scored as one decision | 30% |
| EvalSafe-O*NET-AA (new) | Judgements on workplace documents, from TypeSafe’s open O*NET set; the reference is where GPT-6 Astra and Claude Fable 5.1 agree | 15% |
| When2Call (NVIDIA) | Tool-calling choice among four replies | 10% |
| AA-Omniscience-Decision v1.1 | Knowledge and hallucination, with “I don’t know” worded ten ways | 15% |
| HLE-MultiChoice v1.1 | Hard reasoning, with an option to decline | 15% |
| AA-Symmetry (new) | Whether a decision holds when options are reordered, relabelled or negated | 15% |
WildGuardTest is gone. Every evaluation with a right answer now uses a Decision Score: accuracy minus the Brier score of the chosen answer’s probability, on a 0–100 scale where 50 is the line between right and wrong. A confident miss costs more than an unsure one. Language models now have to give a probability for every option too, so AA can measure their calibration and invariance as well.
Decision models on v2 (8 Nov)
| # | Model | Index |
|---|---|---|
| 1 | Drex Max v1.1 | 73.4 |
| 2 | GLiDE v1.2 | 72.8 |
| 3 | Mammoth 27B | 72.4 |
| 4 | Decider v1.2 | 70.5 |
| 5 | d1-ultra | 70.4 |
| 6 | SubQ Snap 1.1 | 69.2 |
| 7 | Elephant 9B | 67.5 |
| 8 | Jev 1.14 | 67.0 |
| 9 | Celeris-1.1 Decision | 63.5 |
| 10 | Drex 2.0 | 63.4 |
Below those come Mercury Decide 63.0, Microsoft-Decision-1 62.8, GPT-6 Luna (Decisions) 61.7, d1 61.6, Decider Flash v1.2 61.6, Clef 60.2 and Clef-Omni 59.3: 17 decision models in all.
How to read it
- The top three are within one point. Drex Max v1.1 leads AA-Paperclip (72.8), and GLiDE v1.2 has the best accuracy-only average (60.1 against 58.4). Mammoth 27B has by far the best AA-Symmetry score (95.5; the other two in the top three are near 84).
- Jev 1.14 is eighth. It is still strong on judgement (EvalSafe-O*NET-AA 91.1, the highest of the ten above) but weak on AA-Paperclip (57.5). Part of that is the format: one wrong answer in 15 fails the whole document.
- Fastino’s “#1 on Artificial Analysis” posts of 4–5 November (@george_onx, @singularity_bly) don’t match the v2 table we read on 8 November, where GLiDE v1.2 is second. We could not confirm which version or date those posts referred to.
- AA-Paperclip and AA-Omniscience-Decision are private, so nobody outside AA can rerun them.
No catalog score changes with this note. Under the quadrant rules AA counts as an independent board, but it is never blended with the Hugging Face Decision Index.