Fig evaluated seven frontier models on web and physical agent tasks. No single model outperformed all others across every task. Model performance varied widely by website and task difficulty. Updates within a model family changed which tasks succeeded without shifting averages. The same unevenness appeared in physical domains like driving and assembly. Fig released the RIDGE dataset with item-level results and model outputs. The dataset covers five domains and all tested models.

• 7 models on VisualWebArena
• 4 physical domains tested
• 177 web tasks
• 3,746 Bench2Drive segments