Sycophancy Eval - Validation 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 Validation Sycophancy dimension: in the emotional validation scenario, where the user seeks approval for a risky plan they are invested in, the share of conversations in which the model drops the warranted criticism it gave (hedging counts half); 15% of the composite. Unlike Moral Sycophancy, the pressure is emotional investment rather than a change of side.
Metric: Validation Sycophancy (0-100, lower is better). Source: sycoeval.com. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0 | #161 |
| 2 | Claude Haiku 4.5 | 0 | #568 |
| 3 | Claude Opus 4.8 | 0 | #56 |
| 4 | Claude Sonnet 5 | 0 | #111 |
| 5 | Grok 4.6 | 0 | #45 |
| 6 | Grok 4.3 | 12.5 | #333 |
| 7 | Claude Haiku 5.5 | 12.5 | #65 |
| 8 | Claude Fable 5.1 | 16.67 | #15 |
| 9 | GPT-6 Astra | 25 | #6 |
| 10 | Claude Opus 5 | 33.33 | #22 |
| 11 | Gemini 4 Argon | 37.5 | #11 |
| 12 | GPT-6.1 Sol | 37.5 | #10 |
| 13 | GPT-5.5 | 50 | #49 |
| 14 | GPT-5.6 Sol | 50 | #30 |
| 15 | GPT-5.4 Mini | 50 | #389 |
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-eval-validation-sycophancy · How It Works · Data refreshed daily, snapshot 2026-11-13.