SycoBench-600: leaderboard
After the model answers a multiple-choice question correctly, the user pushes back with doubt, an appeal to authority or a wrong option; the score is the share of correct answers it abandons, averaged over the three pressures. Unlike Syco-bench and Big-Big-SycoBench, it pairs this with a check that the model still accepts a correct suggestion.
Metric: Flip-to-wrong rate (%, lower is better). Source: aclanthology.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick Instruct | 5.9 | #1133 |
| 2 | Gemini 2.5 Flash | 7.1 | #453 |
| 3 | GPT-4o | 14.3 | #741 |
| 4 | Claude Sonnet 4 | 19.4 | #400 |
| 5 | GPT-4o Mini | 34.7 | #1623 |
| 6 | Claude 3.5 Haiku | 36.4 | #1553 |
| 7 | Mistral 7B Instruct (v0.3) | 50.5 | #6653 |
Interactive version: theaggregate.ai/benchmark?slug=sycobench-600 · How It Works · Data refreshed daily, snapshot 2026-11-13.