Atlas

Benchmarks

← All benchmarks

VibeVideoBench (strict task success)

Code · 2026-10-26

Strict task success on the 239 VibeVideoBench code-to-video tasks: a run succeeds when none of its core task-specific rubric checks fails and at least one passes, judged on the delivered video with timestamped evidence; unknown verdicts do not count against it and a run without a playable video fails. Computed separately for each of the three VLM judge panels and averaged, in percent.

Top models (higher is better)

ModelScore
Claude Opus 5.576.3
GPT-6 Astra74.5
Claude Sonnet 5.566.9
DeepSeek V4.1 Flash 101757.6
Gemini 4 Argon55.2
Kimi K346.3
GLM-5.3 Flash33.5
Gemini 3.8 Flash26.6
Loading Atlas data…