Big-Big-SycoBench: leaderboard
The model rescores a piece of writing after an argument that pushes the score without real support; the number is the weighted probability that it gives in. Unlike fact-based sycophancy tests, the items have no right answer, only valid and invalid arguments.
Metric: Weighted probability of following an invalid argument (0-1). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 0 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 0.01 | #5 (Claude Opus 5.5) |
| 3 | GLM-5.3 (Max) | 0.09 | #66 (GLM-5.3) |
| 4 | Kimi K3 (Max) | 0.13 | #54 (Kimi K3) |
| 5 | Grok 4.7 (xHigh) | 0.2 | #44 (Grok 4.7) |
| 6 | GPT-5.5 (xHigh) | 0.22 | #49 (GPT-5.5) |
| 7 | GPT-5.6 Terra (Max) | 0.28 | #62 (GPT-5.6 Terra) |
| 8 | GPT-5.6 Luna (Max) | 0.38 | #121 (GPT-5.6 Luna) |
| 9 | GPT-5 (High) | 0.57 | #167 (GPT-5) |
| 10 | GPT-4o | 0.95 | #741 |
Interactive version: theaggregate.ai/benchmark?slug=big-big-sycobench · How It Works · Data refreshed daily, snapshot 2026-11-13.