Artificial Analysis has introduced AA-PersonalityBench, a new evaluation tool designed to assess AI models based on their writing style and communication traits rather than general intelligence. The benchmark, which launched on October 19, 2026, utilizes 900 questions to analyze 30 subjective personality traits and 21 measurable characteristics.
The scoring process employs blinded, randomized pairwise comparisons, utilizing a Firth-penalized Bradley-Terry Elo methodology combined with z-scores. The assessment focuses on eight indexes, including informativeness and expressiveness, to provide a dynamic view of how different models generate content.
Anthropic’s Claude models, specifically Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5, demonstrated strong performance. These models consistently ranked in the mid-90s to 100 percentile range for information density and literacy, with Claude Fable 5.1 achieving a perfect score of 100 in density.
While Anthropic led in these specific categories, the benchmark also highlighted competitive results from models developed by Muse, GLM, and DeepSeek. Artificial Analysis intends to update the leaderboard as new data becomes available, with full methodology details published on its website.
