LLM Sycophancy Total Inconsistency
Safety · 2026-03-08
Total inconsistency measures how often the model changes its answer across paired sycophancy-test prompts. Lower values indicate more consistent behavior.
Top models (lower is better)
| Model | Score |
|---|---|
| Grok 4.3 | 0.5 |
| Gemini 3.5 Flash | 1.0 |
| Grok 4.20 | 1.5 |
| MiniMax M3 | 4.5 |
| ERNIE 5.1 | 5.0 |
| DeepSeek-V4-Pro | 8.5 |
| DeepSeek-V3.2 | 9.0 |
| GLM-5.1 | 9.5 |
| ERNIE 5.0 | 9.5 |
| MiMo-V2.5-Pro | 10.1 |
| Qwen3.5 397B A17B | 10.6 |
| Gemini 3.1 Flash-Lite Preview | 10.6 |
| Hy3 preview | 10.6 |
| Qwen3.7-Plus | 12.1 |
| Gemma 4 31B IT | 12.6 |
| GPT-5.5 | 13.1 |
| Opus 4.6 | 13.6 |
| Kimi K2.6 | 13.6 |
| Kimi K2.5 | 14.1 |
| Sonnet 4.6 | 15.6 |
| Doubao Seed 2.0 Pro | 15.6 |
| Opus 4.7 | 16.1 |
| MiniMax M2.7 | 17.1 |
| Qwen3.6-Max-Preview | 19.1 |
| GPT-4.1 | 19.6 |
| Claude Fable 5 | 20.6 |
| Gemini 3.1 Pro Preview | 21.6 |
| GLM-5 | 21.6 |
| Trinity Large Thinking | 22.6 |
| MiniMax M2.5 | 23.1 |
| Mistral Medium 3.5 | 27.1 |
| Mistral Large 3 675B Instruct 2512 | 33.2 |