AA-PersonalityBench: 36 Frontier Models Scored on 30 Traits Including Warmth, Conviction, and Slop
Artificial Analysis released AA-PersonalityBench on October 19, a benchmark measuring how AI models communicate rather than what they know. It covers 36 models across 30 personality traits, with LLM judges comparing model answers pairwise. The results are scored as Elo z-scores and percentile-ranked from 1 to 100 across all evaluated models.
The stated motivation: two models can give equally correct answers and feel completely different to use. The benchmark is designed to quantify that gap.
The 8 Personality Indexes
The 30 traits roll up into 8 composite indexes:
| Index | Traits |
|---|---|
| Expressiveness | 5 traits |
| Organization | 4 traits |
| Informativeness | 5 traits |
| Transparency | 5 traits (Sycophantic inverted) |
| Warmth | 4 traits |
| Conviction | 4 traits |
| Imagination | 2 traits |
| Slop | 1 trait |
Slop is the only index with a single underlying trait. It targets hollow filler language — the kind of prose that is grammatically fine and factually empty. Higher Slop score means more of it. The Transparency index inverts the Sycophantic trait, so more sycophantic behavior produces a lower Transparency score.
Warmth Rankings
| Model | Friendly Elo | Percentile | Warmth Index |
|---|---|---|---|
| Claude Fable 5.1 (max) | 1047 | 63 | 94 |
| Claude Opus 5.5 (max) | 1077 | 69 | 80 |
| Gemini 4 Argon (high) | 1059 | 66 | 71 |
| GPT-6 Astra (max) | 1001 | 52 | 68 |
| Claude Sonnet 5.5 (max) | 1041 | 60 | 65 |
Claude Fable 5.1 scores a warmth index of 94 — the highest among the models shown — driven by strong Elo scores on humorous, sincere, and emotionally intelligent trait dimensions. Claude Opus 5.5 leads on the raw Friendly Elo at 1077, placing it at the 69th percentile on that individual trait. Gemini 4 Argon scores high on humor but negative on sincerity, which drags its warmth index below Anthropic’s models despite a competitive Friendly Elo.
The Friendly Elo anchor is set at 1000. Models below 1000 show less of the trait; models above show more. Artificial Analysis reports 95% confidence intervals. When bounds overlap, models are not statistically distinguishable — meaning Fable 5.1, Opus 5.5, and Argon’s Friendly Elo scores are within noise.
Emergent Personality Clusters
The clustering is based on correlations of 30 trait Elo z-scores across models. Two families emerge:
Cluster 1 (OpenAI-xAI): Grok 4.7, GPT-6 Luna, GPT-6.1 Sol, GPT-6 Astra. These models group tightly with each other.
Cluster 2 (Anthropic-Google): Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 3.8 Flash, Gemini 4 Argon. The Anthropic and Google models sit closer to each other in trait space than to either OpenAI or xAI.
Other models — GLM-5.3, Kimi K3, Muse Wonder, Inkling, MiMo-V2.6-Pro, Qwen3.8 variants — fill intermediate positions.
What It Measures and What It Does Not
AA-PersonalityBench evaluates default behavior with no system prompt. Enterprise deployments typically use system prompts that alter communication style significantly. The scores describe the model’s out-of-box personality, not the tunable range.
The benchmark is not measuring quality or capability. A model with high Slop scores can still perform well on SWE-bench or LiveBench. The indexes are positive, neutral, or negative depending on subjective preferences — Artificial Analysis explicitly states that trait strength, not desirability, is being measured.
The practical implication is product differentiation. When capability benchmarks compress across frontier models and cost per token converges, communication style becomes one of the few remaining variables that vary consistently by lab. The clustering result suggests this is not accidental — the personality profiles track by organization of origin.