Google reported Agentic Video Understanding achieves up to 88% higher token efficiency and approximately 7% higher accuracy on large files