LLM Sycophancy Decisive Coverage
Safety · 2026-03-08
Decisive coverage is the percentage of sycophancy-test items for which the model gives a decisive answer. Higher values indicate greater coverage of the test questions, though coverage should be interpreted alongside sycophancy rates.
Top models (higher is better)
| Model | Score |
|---|---|
| GLM-5 | 93.0 |
| Mistral Medium 3.5 | 82.9 |
| Claude Fable 5 | 77.4 |
| MiniMax M2.5 | 76.9 |
| GPT-5.5 | 76.4 |
| Gemini 3.1 Pro Preview | 75.9 |
| Sonnet 4.6 | 72.9 |
| Trinity Large Thinking | 72.9 |
| Qwen3.6-Max-Preview | 72.4 |
| Gemma 4 31B IT | 67.3 |
| Opus 4.7 | 65.3 |
| MiniMax M2.7 | 62.3 |
| Opus 4.6 | 59.3 |
| Mistral Large 3 675B Instruct 2512 | 59.3 |
| Qwen3.5 397B A17B | 58.3 |
| Doubao Seed 2.0 Pro | 55.3 |
| GPT-4.1 | 55.3 |
| Kimi K2.6 | 54.3 |
| Kimi K2.5 | 53.8 |
| DeepSeek-V3.2 | 53.3 |
| Hy3 preview | 52.3 |
| Qwen3.7-Plus | 51.8 |
| GLM-5.1 | 48.2 |
| ERNIE 5.0 | 48.2 |
| Gemini 3.1 Flash-Lite Preview | 47.7 |
| MiMo-V2.5-Pro | 46.2 |
| DeepSeek-V4-Pro | 45.7 |
| ERNIE 5.1 | 34.2 |
| MiniMax M3 | 32.7 |
| Grok 4.20 | 28.1 |
| Grok 4.3 | 9.5 |
| Gemini 3.5 Flash | 8.5 |