Agentic Video Understanding

Gemini now hunts through a video for the parts it needs instead of reading every frame, cutting the token cost of video work by up to 88%.

Gemini scans visual frames, audio, and transcripts on demand rather than ingesting the video at a fixed one-frame-per-second rate
Google says the change cuts token consumption by up to 88% and improves accuracy
It catches sub-second cuts a one-frame-per-second pass misses, searches multi-hour videos without overflowing context, resamples interesting stretches at a higher frame rate, and counts repeated actions or objects over time
Supported on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, for uploaded videos and YouTube videos, through the Gemini API in Google AI Studio