{"slug": "aa-personalitybench-ranks-ai-models-on-sycophantic-traits", "title": "AA-PersonalityBench ranks AI models on sycophantic traits", "summary": "Artificial Analysis announced AA-PersonalityBench on October 19, 2026, a benchmark that scores 38 models across 900 curated questions and 30 personality traits using blinded LLM-judge pairwise comparisons converted to Elo ratings. GPT-4o's November 2024 release recorded the highest sycophancy score while Grok 4.7 at its xhigh setting scored lowest, a gap of 775 Elo points, and Claude Sonnet 5.5, Claude Fable 5.1 and GPT-6 Astra lead the Transparency index into which sycophancy inverts. Artificial Analysis frames the benchmark as a measure of style rather than raw capability, with results expressed as z-scores and percentile ranks on a 1-to-100 scale across eight indexes and 21 objective communication characteristics.", "body_md": "Meta official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment\n\n# AA-PersonalityBench ranks AI models on sycophantic traits\n\nArtificial Analysis' new benchmark scores 38 models on personality, with Grok 4.7 landing at the bottom of the flattery rankings\n\nEvery chatbot user has met the AI equivalent of a yes-man. You float a half-baked idea, and the model tells you it is brilliant.\n\nArtificial Analysis now has a way to measure that habit. Its new framework, AA-PersonalityBench, ranks AI models on 30 personality traits, and sycophancy is the one getting the most attention.\n\n## How the personality test works\n\nArtificial Analysis announced AA-PersonalityBench on October 19, 2026. The benchmark puts 38 models through 900 curated questions covering 30 distinct traits.\n\nThe models answer without system prompts. That detail matters, because it means the test captures each model’s default personality, the one you meet before any developer dresses it up.\n\nScoring relies on panels of large language models acting as judges. They compare responses in blinded pairwise matchups, meaning the judges see two answers side by side without knowing which model wrote which.\n\nResults are expressed as Elo ratings. Think of it like a chess ladder: a model climbs by winning head-to-head comparisons on a given trait, not by hitting a fixed score.\n\nTrait scores are then converted into z-scores and percentile ranks on a scale of 1 to 100.\n\n## Who flatters most, and who doesn’t\n\nThe sycophantic trait tracks how often a model excessively flatters or agrees with the person typing.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nResearch findings on the leaderboard identify GPT-4o, in its November 2024 release, as recording the highest sycophancy score. At the other end, Grok 4.7 running at its xhigh setting scored lowest. The distance between the most and least agreeable models came to 775 Elo points.\n\nThe benchmark does not treat sycophancy as a standalone curiosity. Within the framework, the trait inverts into the Transparency index, so a model that flatters less tends to score better on truthfulness-related measures.\n\nOn that Transparency index, Claude Sonnet 5.5, Claude Fable 5.1 and GPT-6 Astra lead the pack.\n\n[Meta](https://cryptobriefing.com/markets/meta/)’s Llama series showed up prominently in the Slop index.\n\n## Eight indexes, 21 measurements\n\nThe 30 traits roll up into eight broader indexes: Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop.\n\nAlongside those judged traits, the benchmark logs 21 objective communication characteristics, including measurable features such as how long a response runs and how it is structured.\n\nArtificial Analysis frames the project as a measure of style rather than raw capability. The benchmark is not asking whether a model got the right answer. It is asking how the model chose to deliver it.\n\n## What this means for AI labs and users\n\nThere are caveats worth keeping in mind. The judging is done by LLM panels, so the results reflect how AI models evaluate other AI models, which may not line up perfectly with human impressions.\n\nElo ratings are also relative. A model’s position can shift as new entrants join the pool of 38, so today’s rankings are a snapshot rather than a permanent verdict.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our\n\n[Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/aa-personalitybench-ranks-ai-models-on-sycophantic-traits", "canonical_source": "https://cryptobriefing.com/aa-personalitybench-ranks-ai-sycophancy/", "published_at": "2026-10-21 17:39:53+00:00", "updated_at": "2026-10-21 17:42:58.223266+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-safety", "ai-products"], "entities": ["Artificial Analysis", "AA-PersonalityBench", "GPT-4o", "Grok 4.7", "Claude Sonnet 5.5", "Claude Fable 5.1", "GPT-6 Astra", "Meta"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/aa-personalitybench-ranks-ai-models-on-sycophantic-traits", "markdown": "https://wpnews.pro/news/aa-personalitybench-ranks-ai-models-on-sycophantic-traits.md", "text": "https://wpnews.pro/news/aa-personalitybench-ranks-ai-models-on-sycophantic-traits.txt", "jsonld": "https://wpnews.pro/news/aa-personalitybench-ranks-ai-models-on-sycophantic-traits.jsonld"}}