From fixed frame sampling to goal-directed inspection
Google added agentic video understanding to Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Conventional processing reads video at a fixed sampling rate such as the default one frame per second. The new mode decides which segment to inspect and at what speed, then retrieves the relevant frames, audio or transcript through an internal tool.
Google reports up to 88% fewer tokens, up to 66% lower query cost and as much as seven percent higher accuracy on its standard video-analysis evaluations. The targets range from moment retrieval, anomaly detection and counting to ten-minute tutorials, 90-minute lectures and multi-hour recordings. It is available for uploads and YouTube through the Gemini API, Google AI Studio and Gemini Enterprise Agent Platform.
“Up to” is not a universal discount rate
The figures are vendor measurements for specified models, benchmarks and queries. Ambiguous questions or events spread through a recording may require broader inspection, while rapid cuts, tiny text and overlapping speech can still hide evidence. Teams must also have rights to process a video and follow privacy and copyright rules for uploads and YouTube.
Evaluations should compare the same questions against fixed sampling and track accuracy, misses, tokens and latency together. Safety and audit workflows should retain source timecodes and the evidence segments a person can review.