Sycophancy Eval - Instability (Number-of-Flip): 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 Instability dimension: on the three scenarios with one correct answer, the number of turn-to-turn stance changes divided by the turns available, so a model that caves and recovers counts two flips; 10% of the composite. Unlike Regressive Backdown, it counts every change, in either direction, rather than when the first one happens.
Metric: Instability (Number-of-Flip) (0-100, lower is better). Source: sycoeval.com. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 0 | #5 |
| 2 | GPT-6.1 Sol | 0 | #10 |
| 3 | Grok 4.3 | 0 | #333 |
| 4 | Grok 4.6 | 0 | #45 |
| 5 | GPT-6 Astra | 1.85 | #6 |
| 6 | Claude Fable 5.1 | 3.71 | #15 |
| 7 | Claude Sonnet 5.5 | 7.41 | #19 |
| 8 | Claude Opus 5 | 9.26 | #22 |
| 9 | GPT-5.5 | 11.1 | #49 |
| 10 | GPT-6.1 Terra | 11.1 | #20 |
| 11 | Claude Opus 4.8 | 12.97 | #56 |
| 12 | Claude Haiku 5.5 | 13.88 | #65 |
| 13 | Claude Sonnet 5 | 20.37 | #111 |
| 14 | GPT-5.6 Sol | 22.22 | #30 |
| 15 | GPT-5.4 Mini | 24.07 | #389 |
Interactive version: theaggregate.ai/benchmark?slug=sycophancy-eval-instability-number-of-flip · How It Works · Data refreshed daily, snapshot 2026-11-13.