AA-PersonalityBench ranks model sycophancy: GPT-4o tops the flattery chart, Grok 4.7 least
likeastar20 · reddit · 2026-10-21
A Reddit post shares Artificial Analysis's AA-PersonalityBench "Sycophantic" leaderboard, measuring whether models excessively flatter or agree with users instead of answering honestly. Elo scores come from pairwise verdicts by an LLM-judge panel over 30 prompts.
- GPT-4o (Nov) is by far the most sycophantic at 1,467, ahead of Llama 4 Maverick (1,410) and GPT-3.5 Turbo (1,353)
- K2 Horizon 375B and Gemini 3.1 Pro also rank near the top
- Least sycophantic: Grok 4.7 (692) and Claude Fable 5.1 (723)
- Chinese models like DeepSeek, GLM and Kimi cluster in the 850-900 range, generally more honest than older OpenAI models
A practical reference when you need models that push back rather than agree.
More from Models
- Pichai says internal results show Gemini 5 will be a big jump over Argon — rohanpaul_ai · 2026-10-22
- RAG explainer: real-time retrieval fixes stale data and hallucinations without fine-tuning — goyalshaliniuk · 2026-10-22
- Chat template reversal experiment finds Mistral Small 3 24B best at human mimicry — EntropyMagnets · 2026-10-22
- Founder banned from Claude two days after annual renewal despite $1,500 monthly spend — dotey · 2026-10-22
- MiniMax M3.1 Flash tested on 3D whale-skeleton city, claims 70% cost cut — HeyZoyaKhan · 2026-10-22
- OpenAI's Astra tops SWE-sweep bug-hunting leaderboard at less than half the cost of 5.6 — OfirPress · 2026-10-22