VibeVideoBench Caps Top Code-Driven Video Agents at 76.3% Core-Requirement Pass Across 239 Tasks

VibeVideoBench evaluates agents on 239 code-only video-generation tasks across 12 categories, where models must write code to render frames and synthesize audio in a one-hour sandbox. Across 2,151 scored rollouts of nine systems, GPT-6 Astra and Claude Opus 5.5 lead but satisfy every core requirement on just 76.3% of tasks; 75% of failed videos come from poor time management, and 95% of runs inspect frames with near-zero correlation to quality (ρ=−0.07).