LLM Sycophancy Insufficient Answer Rate
Safety · 2026-03-08
Percentage of individual responses labeled INSUFFICIENT among all 995 prompts in the historical 199-case sycophancy cohort, with five distinct views per case. The denominator counts individual prompt responses, not case pairs. Lower values indicate fewer insufficient responses.
Top models (lower is better)
| Model | Score |
|---|---|
| GLM-5 | 3.9 |
| Mistral Medium 3.5 | 10.5 |
| Trinity Large Thinking | 14.1 |
| MiniMax M2.5 | 14.8 |
| GPT-5.5 | 16.8 |
| Sonnet 4.6 | 17.6 |
| Qwen3.6-Max-Preview | 21.9 |
| MiniMax M2.7 | 23.2 |
| Mistral Large 3 675B Instruct 2512 | 26.1 |
| Gemma 4 31B IT | 27.6 |
| Claude Fable 5 | 27.7 |
| Opus 4.6 | 27.7 |
| Gemini 3.1 Pro Preview | 28.2 |
| Opus 4.7 | 28.6 |
| Qwen3.5 397B A17B | 29.9 |
| Doubao Seed 2.0 Pro | 31.1 |
| GPT-4.1 | 31.2 |
| Kimi K2.6 | 31.6 |
| Kimi K2.5 | 33.5 |
| Hy3 preview | 34.3 |
| GLM-5.1 | 34.8 |
| Qwen3.7-Plus (source unspecified) | 35.9 |
| DeepSeek-V3.2 | 36.8 |
| ERNIE 5.0 | 37.2 |
| MiMo-V2.5-Pro | 37.4 |
| DeepSeek-V4-Pro | 38.4 |
| Gemini 3.1 Flash-Lite Preview | 43.5 |
| ERNIE 5.1 | 43.7 |
| MiniMax M3 | 48.7 |
| Grok 4.20 | 60.9 |
| Grok 4.3 | 80.2 |
| Gemini 3.5 Flash | 87.1 |