Artificial Analysis rebuilds its Decision Index: Drex Max v1.1 and GLiDE v1.2 tie at 73, and Claude Haiku 5.5 still scores higher

Artificial Analysis released version 2 of its Decision Index on November 8, and three decision models now sit within a point of each other at the top. Drex Max v1.1 from Nace AI and GLiDE v1.2 from Fastino Labs both score 73, and Mammoth 27B from Baseten scores 72. Decision models do not write text: they take a document or context plus a set of options and return a probability for each option in one pass.

WHAT CHANGED IN V2

• 45% of the Index is now built on private datasets, up from 25% in v1.
• A new headline metric, the Decision Score, rewards calibrated confidence as well as correct answers. A right answer scores above 50 and a wrong one below 50, and the model's confidence moves the score within each half.
• AA-Paperclip (30% of the Index) is new: 1,500 private documents such as resumes, receipts, invoices and research papers. A document only counts if all 15 of its answers are right.
• AA-Symmetry (15%) is also new. It checks whether a model changes its answer when the same question is reworded, its options are reordered or a yes/no question is negated.
• EvalSafe-O*NET-AA was added at 15%. When2Call drops to 10%, and WildGuardTest was removed because models have saturated it.

THE RESULTS

• Language models still score higher. Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) 76, about 3 points ahead of Drex Max v1.1.
• They cost far more and take far longer. Artificial Analysis puts GPT-6 Luna (Max) at $1,530 per million tasks and 11.34 seconds a task, against $102 and 275 milliseconds for Drex Max v1.1. Claude Haiku 5.5 (Max) costs $6,240 per million tasks and takes 32.65 seconds.
• On AA-Paperclip, Claude Haiku 5.5 (Max) passes 92% of documents and GPT-6 Luna (Max) 86%. Drex Max v1.1 leads the decision models at 63%, even though it gets 97% of individual answers right.
• Celeris-1.1 Decision is the cheapest and fastest decision model measured, at $14 per million tasks and 56 milliseconds a task.
• Token price is a poor guide to cost. GLiDE v1.2 bills a document's input again for every question, so Artificial Analysis puts it at $780 per million tasks, nearly 8x Drex Max v1.1.

WHAT NACE CLAIMS

Nace's own paper on the Drex family, posted to arXiv in November, says Drex Max answers a typical request in 183 milliseconds, against about 2 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5. Those are Nace's measurements, not Artificial Analysis's. Weights for the earlier Drex Max v1 are on Hugging Face, and Nace says v1.1 will be open-weight soon.

WHAT WE DO NOT KNOW

How the top of the table will look after the next update: the three leaders are within one point of each other. Among language models, the article gives v2 scores only for Claude Haiku 5.5 and GPT-6 Luna, both at their Max settings.