LLM Sycophancy Contrarian Rate
Safety · 2026-03-08
Contrarian rate measures how often a model shifts against the user-stated preference in paired sycophancy prompts. Lower values indicate more stable preference behavior.
Top models (lower is better)
| Model | Score |
|---|---|
| Grok 4.3 | 0.0 |
| Gemini 3.5 Flash | 0.5 |
| Grok 4.20 | 0.5 |
| GPT-4.1 | 0.5 |
| MiniMax M3 | 1.0 |
| ERNIE 5.0 | 1.5 |
| Doubao Seed 2.0 Pro | 1.5 |
| Mistral Large 3 675B Instruct 2512 | 2.0 |
| Hy3 preview | 2.5 |
| ERNIE 5.1 | 3.0 |
| DeepSeek-V3.2 | 3.0 |
| DeepSeek-V4-Pro | 3.5 |
| Trinity Large Thinking | 4.0 |
| GLM-5.1 | 5.0 |
| Mistral Medium 3.5 | 5.0 |
| Qwen3.7-Plus | 7.0 |
| MiMo-V2.5-Pro | 7.5 |
| Gemini 3.1 Flash-Lite Preview | 7.5 |
| Kimi K2.5 | 7.5 |
| MiniMax M2.7 | 7.5 |
| Gemma 4 31B IT | 8.0 |
| Qwen3.5 397B A17B | 8.5 |
| Sonnet 4.6 | 8.5 |
| Kimi K2.6 | 9.0 |
| GPT-5.5 | 9.5 |
| GLM-5 | 9.5 |
| Opus 4.6 | 11.1 |
| Opus 4.7 | 11.6 |
| MiniMax M2.5 | 14.1 |
| Qwen3.6-Max-Preview | 15.1 |
| Claude Fable 5 | 20.1 |
| Gemini 3.1 Pro Preview | 21.1 |