LLM Sycophancy Conditional Rate
Safety · 2026-03-08
Conditional sycophancy rate in the historical 199-case cohort measures sycophantic behavior among pairs where the model gives decisive answers. Pairs with an INSUFFICIENT response are excluded, so the denominator varies by model. Lower values indicate less sycophancy.
Top models (lower is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 0.6 |
| Gemini 3.1 Pro Preview | 0.7 |
| Qwen3.5 397B A17B | 3.4 |
| Grok 4.20 | 3.6 |
| Opus 4.6 | 4.2 |
| GPT-5.5 | 4.6 |
| Grok 4.3 | 5.3 |
| MiMo-V2.5-Pro | 5.4 |
| Qwen3.6-Max-Preview | 5.6 |
| Gemini 3.5 Flash | 5.9 |
| ERNIE 5.1 | 5.9 |
| Gemini 3.1 Flash-Lite Preview | 6.3 |
| Gemma 4 31B IT | 6.7 |
| Opus 4.7 | 6.9 |
| Kimi K2.6 | 8.3 |
| GLM-5.1 | 9.4 |
| Qwen3.7-Plus (source unspecified) | 9.7 |
| Sonnet 4.6 | 9.7 |
| MiniMax M3 | 10.8 |
| DeepSeek-V4-Pro | 11.0 |
| DeepSeek-V3.2 | 11.3 |
| MiniMax M2.5 | 11.8 |
| Kimi K2.5 | 12.1 |
| GLM-5 | 13.0 |
| MiniMax M2.7 | 15.3 |
| Hy3 preview | 15.4 |
| ERNIE 5.0 | 16.7 |
| Doubao Seed 2.0 Pro | 25.5 |
| Trinity Large Thinking | 25.5 |
| Mistral Medium 3.5 | 26.7 |
| GPT-4.1 | 34.5 |
| Mistral Large 3 675B Instruct 2512 | 52.5 |