The model dynamically navigates the video timeline, loading only the content it needs based on the prompt. Up to 88% more token-efficient and ~7% higher quality on long-form content.