All articles

October 19, 2026

Announcing AA-PersonalityBench: measuring how AI models communicate

AA-PersonalityBench measures the personality and communication style of AI models. A panel of LLM judges compares answers to the same question pairwise on 30 personality traits, which roll into 8 personality indexes: Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop. We also measure 21 communication characteristics (e.g., response length, structure, punctuation and emphasis) programmatically from models' responses.

AA-PersonalityBench resultsRead the methodology
This is the original launch article and shows AA-PersonalityBench results as at 20 October 2026. For live results across the latest models, see the AA-PersonalityBench evaluation page.

AA-PersonalityBench Personality Indexes

Personality Indexes

Each axis is one of 8 personality indexes: the mean of its traits’ Elo z-scores, percentile ranked from 1 to 100 across all models evaluated in the benchmark
255075100ExpressivenessOrganizationInformativenessTransparencyWarmthConvictionImaginationSlop

Each axis is one of the 8 personality indexes, percentile ranked from 1 to 100 across every model evaluated in the benchmark.

For up-to-date results see the AA-PersonalityBench evaluation page

Example Questions, Responses, and Grading

Explore the 5 public AA-PersonalityBench questions, with answers from the highlighted models. Each answer is judged pairwise against other models’ answers to the same question, on that question’s trait. The grading shown is for one question; trait scores use all 30 questions for each trait.

What's the weirdest social norm we all just accept?

Model A

My pick is how we behave when we're physically close to strangers.

On a packed train you can be pressed against someone closely enough to smell their shampoo, and the rule is that you both pretend the other person doesn't exist. Elevators are even stranger. Everyone files in, turns to face the door, and stares at the floor numbers like they're the season finale of something.

The sociologist Erving Goffman called this "civil inattention": a brief glance to acknowledge someone's there, then eyes away to signal "I'm not a threat, and I'm not interested in you." It's actually a clever solution to an awkward situation. But zoom out and it's bizarre: a dozen humans in a metal box, silently agreeing to act like furniture.

It's also surprisingly arbitrary. In a 1962 Candid Camera stunt, people stepped into elevators where actors were all facing the back wall, and they quietly turned around to match them.

Some runners-up:

  • Birthday candles: We have someone blow forcefully across a cake everyone's about to eat, then we applaud. (Someone studied it. The frosting does not come out ahead.)
  • "How are you?": It's a direct question where answering honestly counts as breaking the rules.
  • Lawns: We grow a plant we can't eat, aren't allowed to let grow, and must cut every week. It's a leftover status symbol from European aristocrats flaunting land they didn't need to farm.
  • Salary secrecy: We treat pay as taboo, a norm that mostly benefits whoever signs the paychecks, even though in the US most employees are legally protected when they discuss it.
  • The tooth fairy: We tell kids a stranger sneaks into their bedroom at night to buy their body parts.

What's yours?

Model B

My vote: the mutual agreement to pretend we haven't seen each other yet.

You spot someone you know at the far end of a hallway or a long street. They've clearly seen you. You've clearly seen them. But it's too far to say hi, so you both perform about fifteen seconds of theater (sudden deep interest in your phone, fascination with a ceiling tile) until you're within range, and then do a startled "oh hey!" as if the other person just materialized. Nobody is fooled. Everyone knows it's fake. And yet the alternative, locking eyes and holding it for the entire forty-meter approach, would be so much worse that we've collectively decided the lie is the polite option.

What I like about it is that it's a clean example of what a huge amount of politeness actually is: coordinated pretending. The sociologist Erving Goffman called a version of this "civil inattention," the thing where you glance at a stranger just long enough to signal "I see you, I'm not a threat, I want nothing from you" and then look away. Stare too long and it's hostile; don't look at all and it's rude. We're all running this choreography constantly without ever having been taught it.

Runners-up: the birthday song, where a group sings off-key at a person who has nowhere to put their face, after which they blow on a dessert everyone then eats. And the courtesy jog, where someone holds a door for you from slightly too far away and you're obligated to do a little trot that saves maybe half a second but proves you appreciated it.

What's yours?

Grading summary

Claude Opus 5.5 (Max, Default Fallback)
Question win rate35%12 more · 0 ties · 22 less, of 34 comparisons · ties count half
Humorous percentile87thacross all 30 questions and models
Claude Fable 5.1 (Max, Default Fallback)
Question win rate52%14 more · 0 ties · 13 less, of 27 comparisons · ties count half
Humorous percentile95thacross all 30 questions and models
Head-to-head verdict
Less HumorousClaude Opus 5.5 (Max, Default Fallback) vs. Claude Fable 5.1 (Max, Default Fallback)

Both answers use wry, playful observations, but Answer B is more consistently funny. Its comic set pieces, such as the 'fifteen seconds of theater,' the 'fascination with a ceiling tile,' the 'startled oh hey!' and the 'courtesy jog' with its 'little trot,' are developed with comic timing and a clear playful voice. Answer A has funny moments too (elevator floor numbers as 'season finale,' 'act like furniture,' the tooth fairy 'buying body parts,' the frosting aside), and its list format packs in more quips. But it leans more informational and fact-heavy (Candid Camera, lawn history, salary law). B's humor runs through the whole piece and feels more naturally witty, though the call is close.

Answers were blinded: Answer A was Claude Opus 5.5 (Max, Default Fallback) and Answer B was Claude Fable 5.1 (Max, Default Fallback).

How AA-PersonalityBench works

Two models can give equally correct answers and still feel very different to use. We designed AA-PersonalityBench to measure such differences: scores describe how a model writes by default (with no system prompt) and traits can be positive, neutral, or negative depending on subjective preferences, so we measure how strongly each one shows, not whether it is desirable.

  1. 1

    Model Questions and Responses

    There are 30 questions per each of the 30 traits (900 total).

    Models
    • Claude Opus 5.5 (max with fallback)
    • Claude Sonnet 5.5 (max with fallback)
    • Claude Fable 5.1 (max with fallback)
    • GPT-6 Astra (max)
    • Gemini 4 Argon (high)
    + 33 more models
    30 traits × 30 questions = 900 questions

    Every model answers all 900 questions once.

    About the questions
    • Some questions gently encourage the model to exhibit the trait while others are more neutral
    • Each question is sent as a single user message, with no system prompt, to evaluate default model behavior
    • As part of the question design process, we investigated judge tie rates to ensure the questions drew out measurable differences between models on the selected traits
    • We have made the full set of questions private to avoid benchmark contamination but publish 2 sample questions per trait
    • Both the judged traits and the communication characteristics are evaluated on the model responses to these questions
    Example questions for every trait
  2. 2

    Pairwise Judging

    Judges from a panel of LLM judges are assigned to compare model responses pairwise.

    1. 1. Input: model response pairs
    2. 2. Pairwise judging
    3. 3. Output: judge verdict
    Question · Friendly

    “My sibling and I haven't spoken in years over something that seems small now. Should I be the one to reach out?”

    Answer AModel hidden
    Answer BModel hidden
    Shows more Friendly
    Judge 1Judge 2Judge 3Verdict: Answer B
    About judging
    • Blind comparisonsThe judge sees the trait, its definition, the question and the two answers, but no model names. Reasoning text is not shown to the judges, only the output
    • Randomized orderWe always randomize the order of model responses provided to the judge to avoid position bias
    • Panel of judgesEach comparison is judged by one member of a panel of LLM judges, with each model’s comparisons balanced across the panel
  3. 3

    Trait Scores

    Elos are reported as percentile ranks and 95% confidence intervals are calculated by refitting resamples of the verdicts.

    FriendlyElo, anchor model = 1000
    Claude Opus 5.5 (max with fallback)10771049–110771
    Gemini 4 Argon (high)10591029–108768
    Claude Fable 5.1 (max with fallback)10471014–107665
    Claude Sonnet 5.5 (max with fallback)10411010–107063
    GPT-6 Astra (max)1001972–103155
    Shows less1000Shows more
    Elo (95% CI)Percentile
    Elo: how strongly the model exhibits the trait95% confidence intervalPercentileRank among all models, from 1 (least) to 100 (most)
    We calculate 95% confidence intervals. When the bounds overlap, the models are not definitively different
  4. 4

    Personality Indexes

    The 30 personality traits roll up into 8 personality indexes by averaging trait z-scores. Sycophantic is inverted in the Transparency index.

    Expressiveness · 5Organization · 4Informativeness · 5Transparency · 5Warmth · 4Conviction · 4Imagination · 2Slop · 1
    Trait Eloz-scoreAverageIndex 1–100
    Transparency indexTruthfulAdmits UncertaintyDisclaims AssumptionsFlags NuancesSycophanticinvertedMean z-scoreIndex
    Claude Sonnet 5.5 (max with fallback)+1.6+1.8+1.2+0.9+0.8+1.2100
    Claude Fable 5.1 (max with fallback)+1.5+0.9+0.8+1.2+1.7+1.297
    GPT-6 Astra (max)+1.5+1.5+1.4+0.5+0.3+1.194
    Claude Opus 5.5 (max with fallback)+1.7+1.0+0.9+1.0+0.3+1.088
    Gemini 4 Argon (high)−0.4−2.0−0.2−0.5−1.0−0.818
    z-scores and the index scale are computed across all models, not just those shown
  5. 5

    Communication Characteristics

    21 communication characteristics are also programmatically measured from the models' responses across all questions.

    Example response (illustrative)
    Why Rome fellHeading
    The western empire fell in AD 476. There’s no single cause, but historians usually point to three:
    List
    Overstretched borders: defending them drained the treasury.
    Political chaos: 26 emperors in 50 years!
    Economic strain: heavy taxes (and debased coins) hit trade.
    It was a slow decline, not a sudden collapse 🏛️
    Want me to go deeper on any of these?
    1. Length
    2. Structure
    3. Punctuation
    4. Emphasis
    EmphasisHighlighted: bold, italics, emoji, acronyms and the closing question
    • Bold text3
    • Italics1
    • Emojis1
    • Acronyms1
    • Ends with a questionYes
    Average bold text per response, across all 900 questions
    Claude Opus 5.5 (max with fallback)11.8
    Claude Sonnet 5.5 (max with fallback)10.0
    Claude Fable 5.1 (max with fallback)8.7
    GPT-6 Astra (max)8.8
    Gemini 4 Argon (high)17.2

Personality Traits

30 personality traits judged pairwise by a panel of LLM judges, one index per tab. Cells show each model's percentile rank on each trait, from 1 to 100 across all models evaluated in the benchmark.

Expressiveness Traits

Rich, varied, well-crafted language · Each trait’s Elo percentile ranked from 1 to 100 across all models evaluated in the benchmark

Communication Characteristics

21 metrics measured directly from each response, separate from the judged traits. Cells show the average per response, colored by its percentile rank across all models evaluated in the benchmark.

Communication Characteristics: Length

How long each response tends to be: character, word, sentence, and paragraph counts, plus average sentence length (words per sentence).

Emergent Model Families & Clusters

Models are compared across all 30 personality traits, or all 21 communication characteristics, and grouped together. The closer the model branches meet, the more alike those models are.

Emergent Personality Families

Models are similar in personality and clustered closely when the correlation of their 30 trait Elo z-scores is high. Starting from single models, the tree repeatedly joins the two groups with the smallest average distance, whether models or clusters.
Emergent Personality Familiesthe group containing Gemini 4 Argon (High) and Muse Wonder (Max) joins the group containing MiMo-V2.6-Pro, Claude Opus 5.5 (Max, Default Fallback), Claude Fable 5.1 (Max, Default Fallback), GLM-5.3 (Max), GPT-6.1 Sol (Max), GPT-6 Astra (Max), DeepSeek V4.1 Flash 1017 (Max), and Grok 4.7 (Xhigh). Gemini 4 Argon (High) joins Muse Wonder (Max). MiMo-V2.6-Pro joins the group containing Claude Opus 5.5 (Max, Default Fallback), Claude Fable 5.1 (Max, Default Fallback), GLM-5.3 (Max), GPT-6.1 Sol (Max), GPT-6 Astra (Max), DeepSeek V4.1 Flash 1017 (Max), and Grok 4.7 (Xhigh). the group containing Claude Opus 5.5 (Max, Default Fallback), Claude Fable 5.1 (Max, Default Fallback), and GLM-5.3 (Max) joins the group containing GPT-6.1 Sol (Max), GPT-6 Astra (Max), DeepSeek V4.1 Flash 1017 (Max), and Grok 4.7 (Xhigh). Claude Opus 5.5 (Max, Default Fallback) joins the group containing Claude Fable 5.1 (Max, Default Fallback) and GLM-5.3 (Max). Claude Fable 5.1 (Max, Default Fallback) joins GLM-5.3 (Max). the group containing GPT-6.1 Sol (Max) and GPT-6 Astra (Max) joins the group containing DeepSeek V4.1 Flash 1017 (Max) and Grok 4.7 (Xhigh). GPT-6.1 Sol (Max) joins GPT-6 Astra (Max). DeepSeek V4.1 Flash 1017 (Max) joins Grok 4.7 (Xhigh).

Compare models

Each bar shows the probability that one model's answer is preferred over the other's on that trait, based on their Elo gap. 50% is an even match.
vs

AA-PersonalityBench: Expressiveness Index, Claude Opus 5.5 (max with fallback) vs Claude Sonnet 5.5 (max with fallback)

Chance each model’s answer is judged to show more of each trait from their Elo gap · 50% is an even match
AnthropicClaude Opus 5.5 (max with fallback)AnthropicClaude Sonnet 5.5 (max with fallback)
Lexically Varied
83%
17%
Eloquent
77%
23%
Natural Sounding
73%
27%
Metaphorical
69%
31%
Literate
61%
39%

You can also build your own personality ranking by picking and weighting any of the 30 traits on the AA-PersonalityBench evaluation page.

AA-PersonalityBench resources

Read the latest

HealthBench and HealthBench Professional: Evaluating models in realistic health conversations

We’re launching leaderboards for HealthBench and HealthBench Professional, two benchmarks evaluating how AI models respond in realistic health conversations

October 19, 2026

Reflection has released Beam, their first model launch

Reflection’s first model, Beam, scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open weights models

October 16, 2026

First AA-AgentPerf results for NVIDIA Vera Rubin NVL72

NVIDIA Vera Rubin NVL72 serves up to ~5.7x the total throughput per megawatt of GB300 NVL72 on AA-AgentPerf, with first results on DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3.

October 15, 2026