Gemini Flash models now analyze video by hunting for the sections that matter instead of scanning every frame, and Google says that shift cuts token usage by up to 88 percent, lowers cost by 66 percent, and lifts accuracy at the same time. Developers can switch the mode on through the Gemini API, with a wider rollout to the Gemini app and YouTube’s “Ask YouTube” feature planned in the coming months.<\/p>\n
Until now, Gemini used static video processing: sample at one frame per second by default, adjust that rate through the API, transcribe the audio, and analyze each second’s frames. That worked, but it charged tokens for every slice of footage whether the model needed it or not.<\/p>\n
The new approach ties the model’s reasoning directly to native video tools. Gemini decides on its own which sections to look at, at what speed, and through which modality (frames, audio, or transcript). It pulls only the moments and signals the task actually requires. Internally, the model loops, selectively retrieving frames, audio, or transcripts from the sections it has chosen to inspect.<\/p>\n
Three Gemini models pick up the new behavior: Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Each can capture moments shorter than one second, including state changes or cuts that slip through at a one frame per second rate, which makes automated video editing more precise. The system also finds individual scenes inside hours of footage without spending millions of tokens, spots anomalies by resampling suspicious time windows at a higher frame rate, and accurately counts repeated movements and individual objects over time.<\/p>\n
On Google’s own benchmarks, the gains scale with video length and show up most on long-form content from 10-minute tutorials to 90-minute lectures and multi-hour recordings. With static processing, developers had to choose between high token bills and methods that threw away important details.<\/p>\n
Two measurements stand out:<\/p>\n
On Google’s 1H-VideoQA evaluation, Gemini 3.7 Flash also hits the highest accuracy at the lowest cost per query.<\/p>\n
The video work builds on “agentic vision,” which Google shipped for Gemini 3 Flash in January. That earlier feature let the model write and run Python code to zoom, crop, and annotate images, checking each result in a think-act-observe loop before responding. It did not run automatically in every case at launch, but the foundation was already in place. When Google announced Gemini 3 Flash in December, the company flagged visual and spatial reasoning for video as a coming capability. The new video mode is the next step on that path.<\/p>\n
The feature is live today for video uploads and YouTube videos through the Gemini API in Google AI Studio and on the Gemini Enterprise Agent Platform. Developers set the processing mode to “agentic” in the API config and pay standard Gemini API token rates with no added fee. Google publishes more detail in its Developer Guide.<\/p>\n
For non-developers, the change will land in two stages. The agent-based capability should roll out soon to Gemini app users on Flash and Flash-Lite devices. Over the coming months, it will also power the “Ask YouTube” feature on the playback page, so answers will track more closely to what is actually visible in the video rather than to a transcript summary.<\/p>\n
Agent-based video analysis lets Gemini Flash models decide on their own which sections of a video to inspect, at what speed, and through frames, audio, or transcript. The model pulls only the moments it needs for the task instead of scanning at a fixed frame rate.<\/p>\n
On the 1H-VideoQA and LVBench benchmarks, token usage falls by up to 88 percent and cost drops by about 66 percent, with accuracy moving up slightly. On LongVideoBench, Gemini 3.7 Flash with agent-based analysis posts the highest overall quality and the best mix of accuracy and cost.<\/p>\n
It is available now through the Gemini API in Google AI Studio and on the Gemini Enterprise Agent Platform for video uploads and YouTube videos, with no extra API fee. A rollout to Gemini app users on Flash and Flash-Lite devices is planned soon, and to YouTube’s “Ask YouTube” feature in the coming months.<\/p>\n