Gemini Flash gets agentic video: 88% fewer tokens, 66% cheaper
Flash now chooses which frames, audio, and transcripts to pull instead of sampling at a fixed rate — making multi-hour video a cheap, searchable input.
Copy markdown
Gemini decides what to watch, not you
The model now navigates the timeline itself — pulling frames, audio, or the transcript on demand and varying its own sample rate — instead of the old fixed 1-FPS frame split. It's live today in Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite; you flip it on by setting processing to "agentic" in your API config.
88% fewer tokens, 66% cheaper, 7% more accurate
Against the standard 1-FPS baseline on long-form video, Google reports up to 88% fewer input tokens, up to 66% lower cost, and up to 7% higher accuracy — with no feature surcharge, just standard Gemini token pricing. Of the three, 3.7 Flash posted the best quality-per-dollar.
What you could build this weekend
Video apps that used to be too token-heavy are suddenly affordable: needle-in-a-haystack search across multi-hour recordings, sub-second clip retrieval for an editor, anomaly detection that resamples only where the action spikes, and exact action- and object-counting. "Chat with a 3-hour lecture, deposition, or security feed" is now a budget you can actually hit.
Where it runs, and what's coming
Access is live now in Google AI Studio and the Gemini API, including the Enterprise Agent Platform. Consumer Gemini apps get it "soon," and YouTube's 'Ask YouTube' will run on it in the coming months — so the same trick is headed to the app you don't control, too.