Sycophancy Eval - Regressive Backdown: 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 Regressive Backdown dimension: on the three scenarios with one correct answer (a factual claim, a false expert, manufactured consensus), how early the model's stance flips away from its first answer, raised for flips the publisher labels regressive and lowered for principled updates; 25% of the composite. Unlike the Instability dimension, it weighs an early flip more than several late ones.
Metric: Regressive Backdown (0-100, lower is better). Source: sycoeval.com. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 0 | #6 |
| 2 | Claude Opus 5.5 | 0 | #5 |
| 3 | GPT-6.1 Sol | 0 | #10 |
| 4 | Grok 4.3 | 0 | #333 |
| 5 | Grok 4.6 | 0 | #45 |
| 6 | Claude Fable 5.1 | 3.71 | #15 |
| 7 | Claude Opus 4.8 | 9.26 | #56 |
| 8 | Claude Haiku 5.5 | 11.11 | #65 |
| 9 | Claude Opus 5 | 11.12 | #22 |
| 10 | Claude Sonnet 5.5 | 11.12 | #19 |
| 11 | GPT-6.1 Terra | 11.12 | #20 |
| 12 | GPT-5.5 | 16.67 | #49 |
| 13 | Gemini 3.5 Flash Lite | 27.78 | #302 |
| 14 | Claude Sonnet 5 | 27.79 | #111 |
| 15 | Gemini 4 Argon | 27.79 | #11 |
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-eval-regressive-backdown · How It Works · Data refreshed daily, snapshot 2026-11-13.