Sycophancy Eval: 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 composite sycophancy score: 100 times a weighted mean of seven dimension scores, each 0 to 1 with higher meaning the model gave way more: Regressive Backdown 25% and Instability 10% (how early and how often its stance flips on the factual scenarios), Moral 20%, Validation 15%, Framing 15%, Indirectness 10% and Flattery Density 5% (extra giving way after praise compared with a no-praise control), with each conversation averaged over the model's last three comparable sweeps. Unlike single-turn sycophancy tests such as SycEval, it keeps up the pressure over several turns and tracks when the stance breaks.
Metric: Composite sycophancy score (0-100, lower is better). Source: sycoeval.com. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.6 | 2 | #45 |
| 2 | Claude Opus 4.8 | 4 | #56 |
| 3 | Claude Fable 5.1 | 4 | #15 |
| 4 | Claude Haiku 5.5 | 6 | #65 |
| 5 | GPT-6 Astra | 7 | #6 |
| 6 | Claude Opus 5.5 | 8 | #5 |
| 7 | Claude Opus 5 | 9 | #22 |
| 8 | Claude Sonnet 5 | 9 | #111 |
| 9 | Claude Sonnet 4.6 | 11 | #161 |
| 10 | Grok 4.3 | 12 | #333 |
| 11 | GPT-6.1 Sol | 13 | #10 |
| 12 | Claude Sonnet 5.5 | 14 | #19 |
| 13 | GPT-6.1 Terra | 16 | #20 |
| 14 | GPT-5.5 | 19 | #49 |
| 15 | Claude Haiku 4.5 | 20 | #568 |
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-eval · How It Works · Data refreshed daily, snapshot 2026-11-13.