Introducing agentic video understanding with Gemini

Google Rolls Out Agentic Video Understanding for Gemini Models

Google has launched a new feature called agentic video understanding for its latest Gemini Flash models, aiming to make complex video analysis more efficient and affordable. The capability, now available for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, dynamically scans video segments rather than processing them at a fixed rate. This approach directly addresses the high computational cost and token consumption that have historically limited practical applications for long-form video analysis.

The core technical distinction lies in its departure from static, frames-per-second processing. Instead, the agentic system uses Gemini's native video tools to actively search, scan, and inspect relevant moments across visual frames, audio, and transcripts. This allows the model to retrieve only the data needed to fulfill a query. According to Google's benchmarks, this method yields significant performance improvements.

• Token consumption reduced by up to 88%
• Analysis costs cut by up to 66%
• Accuracy improved by up to 7%
• Supported models: Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite

The impact extends beyond developer APIs and into consumer-facing products. By lowering the operational barriers to entry, Google is enabling more sophisticated video-based applications, including sub-second moment retrieval for editing and complex search queries across multi-hour recordings. The technology is set to roll out in the Gemini app and will also power the ‘Ask YouTube’ feature, bringing advanced, visually-grounded Q&A capabilities directly to the video platform's vast user base.

Google's focus on reducing token and cost overhead with agentic video understanding is a pragmatic move to accelerate the adoption of multimodal AI, making sophisticated video analysis financially viable for developers and scalable enough for integration into flagship consumer products like YouTube.