Artificial Analysis has launched AA-PersonalityBench to compare how models communicate, not whether their answers are correct. The public page describes 30 personality traits grouped into eight indexes and 21 directly measured communication characteristics.
Pairwise LLM-judge comparisons yield Elo ratings for traits; normalized trait scores are aggregated into eight indexes, while 21 response characteristics are measured directly. The source says each trait uses 30 prompts across 30 traits, 900 total, and that verdicts become Bradley-Terry Elo ratings. This turns style into a reported evaluation dimension rather than an informal impression. Model selection increasingly includes user-facing interaction quality alongside task capability, creating demand for comparable style measurements.
Teams can include communication preferences as one model-selection dimension, without treating style scores as a proxy for task performance. The public description gives the benchmark method and displayed counts, but does not establish predictive validity for user outcomes or stability across contexts; its displayed count differs from the supplied launch description. The first watch condition is whether Artificial Analysis clarifies the model count and publishes evidence that style rankings generalize across tasks and settings. The public page displays “10 of 38 models,” while the supplied launch description says 36 LLMs; that discrepancy merits clarification, not speculation about its cause.