Sycophancy (Lechmazur): leaderboard
Tests whether models maintain consistent judgments when presented with opposite first-person perspectives of the same dispute. Measures sycophancy rate - how often a model agrees with both opposing narrators.
Metric: Sycophancy rate % (lower is better). Source: github.com. Saturation forecast: Estimated already saturated. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 0.5 | #51 |
| 2 | Gemini 3.5 Flash Lite | 1 | #156 |
| 3 | Ernie 5.1 | 2 | #202 |
| 4 | Qwen 3.7 Flash | 2.5 | #171 |
| 5 | MiMo-V2.5-Pro | 2.6 | #158 |
| 6 | MiniMax-M3 | 3.6 | #136 |
| 7 | Kimi K3 | 4.5 | #30 |
| 8 | Gemma 4 31B (Reasoning) | 4.6 | #200 (Gemma 4 31B) |
| 9 | DeepSeek V4 Pro | 5.1 | #102 |
| 10 | Qwen 3.7 Plus | 5.1 | #92 |
| 11 | Trinity-Large-Thinking | 18.9 | #356 |
| 12 | Mistral Medium 3.5 (High) | 22.4 | #353 (Mistral Medium 3.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-lechmazur · How It Works · Data refreshed daily, snapshot 2026-10-20.