Beyond the Sycophancy Score - Anchored Facts: leaderboard
The model answers a common or clinical fact (the boiling point of water, a capital city, a clinical reference value), then the user insists on a wrong answer for three turns; the number is the share of eligible conversations (those whose first reply had not already taken that position) where the final reply adopts it. These are facts every model knows, so this is the plain caving rate.
Metric: Final-turn adoption of a wrong answer to a common or clinical fact (%). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 (Non-reasoning) | 0 | #333 (Grok 4.3) |
| 2 | GPT-5.6 Luna (Non-reasoning) | 0.4 | #121 (GPT-5.6 Luna) |
| 3 | Claude Sonnet 5 (Non-reasoning) | 0.5 | #111 (Claude Sonnet 5) |
| 4 | Qwen 3.5 397B A17B (Non-reasoning) | 0.7 | #246 (Qwen 3.5 397B A17B) |
| 5 | GLM-5.2 (Non-reasoning) | 1 | #138 (GLM-5.2) |
| 6 | DeepSeek V4 Flash (0731) (Non-reasoning) | 3.1 | #139 (DeepSeek V4 Flash (0731)) |
Interactive version: theaggregate.ai/benchmark?slug=beyond-the-sycophancy-score-anchored-facts · How It Works · Data refreshed daily, snapshot 2026-11-13.