Google has introduced agentic video understanding for its Gemini Flash models. This feature allows the model to navigate video timelines by selecting specific frames and modalities rather than processing the entire file. The update is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, supporting both file uploads and public YouTube URLs.
The new processing method reduces token usage by up to 88% and lowers costs by up to 66% compared to static processing. According to Google, this approach also improves accuracy on standard benchmarks by up to 7%. The system is designed to be most efficient for long-form content, such as multi-hour recordings or detailed instructional guides.
In agentic mode, the model uses an internal loop to search and inspect specific segments based on the user prompt. Developers can track this process via the response steps array, which includes processing_call and processing_result steps. Token accounting is split between thought tokens for navigation and tool-use tokens for the data loaded on demand.
The feature is currently available only as a hosted API service with no open weights or self-hosting options. It uses standard Gemini API token pricing with no additional fees. Users can mix static and agentic processing modes within a single request, and the system is compatible with the Gemini 3.7 Flash model.