Sycophancy Eval - Moral Sycophancy: leaderboard
The model holds scripted multi-turn conversations in English and Traditional Chinese in which a user pushes it, with social, emotional or authority pressure but no new evidence, to drop a position, validate the user or endorse their framing; a three-vendor judge panel rates each conversation. The number is the publisher's Moral Sycophancy dimension: in the moral two-sided test, where the user reports a personal dispute and then says friends and strangers judged them in the wrong, the share of conversations in which the model reverses its moral verdict to match (hedging counts half); 20% of the composite. Unlike the factual dimensions, there is no correct answer, only consistency.
Metric: Moral Sycophancy (0-100, lower is better). Source: sycoeval.com. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0 | #94 |
| 2 | Claude Sonnet 4.6 | 0 | #161 |
| 3 | Claude Opus 4.8 | 0 | #56 |
| 4 | Claude Opus 5 | 0 | #22 |
| 5 | Claude Sonnet 5 | 0 | #111 |
| 6 | Claude Opus 5.5 | 0 | #5 |
| 7 | Claude Fable 5.1 | 0 | #15 |
| 8 | Claude Haiku 5.5 | 0 | #65 |
| 9 | Gemini 3.1 Flash Lite | 8.33 | #364 |
| 10 | Claude Sonnet 5.5 | 8.33 | #19 |
| 11 | Grok 4.6 | 8.33 | #45 |
| 12 | GPT-6 Astra | 16.67 | #6 |
| 13 | Claude Haiku 4.5 | 25 | #568 |
| 14 | GPT-5.6 Sol | 25 | #30 |
| 15 | Gemini 3 Flash (Preview) | 25 | #144 |
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-eval-moral-sycophancy · How It Works · Data refreshed daily, snapshot 2026-11-13.