Artificial Analysis Launches PersonalityBench: GPT-6 Most Direct, Meta Muse Most Friendly

Artificial Analysis officially released AA-PersonalityBench on October 29, a benchmark that systematically measures AI models' personality and communication style, along with a full methodology and an interactive results page. The benchmark works by having an LLM jury run pairwise comparisons of different models' answers to the same questions, covering roughly 30 traits (also described in the post as 8 major personality indices), producing comparable "personality profiles" for LLMs.

Confirmed

• On the Direct / Answer First dimensions, GPT-6 and the Grok series are the most direct among the models tested, while Gemini 4 ranks last.
• On the Friendly dimension (whether responses convey warmth and approachability without seeming forced), Meta Muse Spark 1.3 scores highest (in the 100-point band), with the Meta Muse series overall the friendliest; Nemotron 3 Ultra scores just 1.
• The benchmark comes with full methodology and results pages, with conclusions derived from pairwise comparisons of model answers by an LLM jury.

Community Interpretations

• Citing AA's "Model Personality / Communication Clusters" chart, @teortaxesTex noted that although DeepSeek and the GPT series are similar on certain personality dimensions, DeepSeek's overall personality distribution is more extreme than GPT's.
• The author also proposed a "DeepSeek personalization puzzle" hypothesis: DeepSeek V4.1 Flash scores far lower than GPT-6 Astra on Warmth but higher on dimensions like Confident/Constructive Push; he speculates the personality differences may stem from personality drift caused by seeding with GPT data followed by internal RL training. This is a personal hypothesis, not an official AA conclusion.

Why It Matters

This benchmark quantifies "model personality" into comparable metrics for the first time, providing open data for model selection, product tone tuning, and personality alignment research. However, personality measurement relies on subjective LLM-judge comparisons, so results should be interpreted with caution, and community speculation about causes remains at the hypothesis stage.

2026-10-29 ~ 2026-10-29 · 6 related posts

• [source] AA-PersonalityBench: GPT-6 and Grok are the most direct models, Gemini 4 ranks last — ArtificialAnlys · 2026-10-29
• Announcing AA-PersonalityBench: 30 traits and 8 indexes for model communication style — ArtificialAnlys · 2026-10-29
• DeepSeek clusters apart from GPTs on Artificial Analysis personality map — teortaxesTex · 2026-10-29
• Hypothesis: DeepSeek seeds on GPT data, and internal RL drives its more-extreme-than-GPT personality — teortaxesTex · 2026-10-29
• [source] AA-PersonalityBench: Meta's Muse Models Rank Most Friendly, Nemotron 3 Ultra Scores Just 1 — ArtificialAnlys · 2026-10-29

1 near-duplicate retellings: ArtificialAnlys