Gemini Flash now navigates video itself, cutting tokens up to 88%
Set it to 'agentic' and 3.7/3.6 Flash and 3.5 Flash-Lite choose which frames, audio, or transcript to pull — 66% cheaper, on the API and YouTube URLs today.
Copy markdown
The video watches itself
Flip an API flag to `agentic` and the model takes a goal-directed role — deciding what to watch, at what speed, and through which modality (frames, audio, or transcript) instead of ingesting every frame. Google reports up to 88% fewer tokens, up to 66% lower cost, and up to 7% higher accuracy versus fixed-rate processing, all at standard token pricing with no surcharge.
Three Flash models, one flag
It's live across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite through Google AI Studio, the standard Gemini API (both uploaded files and YouTube URLs), and the Gemini Enterprise Agent Platform. Set processing to `agentic` in your API config — Google ships code samples and docs to wire it up.
What you can ship this weekend
Point it at a multi-hour recording and get sub-second moment retrieval, reliable action and object counting, and anomaly detection without torching your token budget — long-video agents that were previously too costly to run. Ask for the exact timestamp something happens in a three-hour video and pay only for the seconds it actually inspects.
Heads-up: omni Flash preview retires Sep 30
Separately, if you still call `gemini-omni-flash-preview`, migrate to the GA `gemini-omni-1.1-flash` before September 30, 2026, when the preview endpoint is deprecated. The GA build adds video extension, first/last-frame interpolation, and 360p–4K resolution control.