For years, AI benchmarks have mostly asked one question: did the model get it right? Artificial Analysis now wants to know something else: how did the model sound while getting it right?
The firm’s new AA-PersonalityBench grades AI models on writing style and communication traits. Anthropic’s Claude models came out on top in two categories, Information Density and Literacy.
What the benchmark actually measures
AA-PersonalityBench launched on October 19, 2026, with data available as of October 20. It is built around 900 questions.
Those questions cover 30 subjective personality traits. The traits are grouped into eight indexes, including Informativeness and Expressiveness.
The benchmark also tracks 21 characteristics that can be measured directly. These include things like response length and structure.
Scoring relies on blinded, randomized pairwise comparisons. Two responses are placed side by side with the model names hidden, and one is judged better on a given trait.
Those head-to-head results feed into a Firth-penalized Bradley-Terry Elo scoring methodology, combined with z-scores.
How Claude scored
Anthropic’s lineup performed strongly across the two headline categories. Four Claude models were singled out: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5.
These models consistently landed within the mid-90s to 100 percentile range in Information Density and Literacy.
Claude Fable 5.1 posted a score of 100 in Information Density and 95 in Literacy.
Claude Opus 5.5 scored 97 in both Information Density and Literacy.
Anthropic did not sweep every category. Other high performers included models from Muse, GLM, and DeepSeek, which showed varied scores across the different indexes.
Why style is becoming a scoreboard
AA-PersonalityBench is explicitly focused on stylistic distinctions rather than overall capability. It does not claim to tell you which model is smartest.
The benchmark is also designed to be dynamic. Artificial Analysis can update it as new evaluations come in, and full methodologies and leaderboards are published on artificialanalysis.ai.
What this means for developers, users, and AI companies
For Anthropic, the results reinforce a reputation for clear, substantive writing. Strong showings in density and literacy across four models, from the larger Opus to the smaller Haiku, suggest the trait is consistent across the lineup rather than limited to one flagship.
The research behind the launch argues that firms prioritizing user interaction quality could attract more users, which could in turn affect their market position and valuation.
Blinding and statistical corrections help, but readers should check how the comparisons are judged before treating any score as gospel. The published methodology on artificialanalysis.ai is the place to start.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.