Google Cuts Long-Form Video Costs with Agentic Video Understanding Integration
Google has integrated its Agentic Video Understanding feature into Gemini, its conversational AI model. This enables autonomous video analysis and reduces costs for long-form videos by up to 66%. The new feature allows Gemini to determine which segments of a video are relevant to the query and iteratively re-examine transcripts and frames to construct an answer.
The Agentic Video Understanding feature is available on three models: Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite. Developers can activate it by specifying 'agentic' as the video processing method in the Gemini API with no additional fees.
Google's conventional approach to video input for summarization and Q&A extracts one frame per second from a video and feeds the combined visual and audio data to the AI in a single pass. This is manageable for short clips but consumes enormous amounts of tokens when dealing with long videos, risking missed fleeting events.
The new Agentic Video Understanding approach rethinks this design by dynamically adjusting the frame rate and resolution of target segments and re-reading them as needed. Google has published results comparing conventional static processing with Agentic Video Understanding across three video understanding benchmarks, showing a reduction in token consumption by up to 88% while improving accuracy by up to approximately 7%.