Google cuts Gemini video-analysis costs by letting Flash models skip frames

Google has begun rolling out a new way for its Gemini Flash models to analyze video, one that lets the model choose what to watch instead of processing an entire clip at a fixed rate. The company says the change, called agentic video understanding, cuts token use by up to 88%, lowers cost by up to 66%, and raises accuracy by up to 7% on standard video benchmarks.

Until now, every Gemini model processed video the same way regardless of the question being asked. It pulled one frame per second from the whole timeline, compressed audio to 1 kilobit per second in a single channel, and stamped every second with a timestamp. A model asked to summarize a 90-minute lecture and a model asked to find the exact moment a speaker switches slides did the same amount of work, because the pipeline had no way to skip ahead.

Agentic processing changes that pipeline into a loop. The model reasons about the prompt, then uses built-in video tools to search, scan, and inspect the segments, frames, audio, or transcript passages it actually needs, loading only that material into context. Developers could previously build something similar themselves by chunking video and querying pieces of it, but Google says the new mode runs that loop inside Gemini itself, removing the engineering work of assembling it.

Where the savings show up

The efficiency gains are concentrated in longer video, according to Google, ranging from ten-minute how-to guides to recordings that run several hours. Short clips see less benefit: Google says static processing, the original fixed-rate method, remains the better choice for videos under five minutes or for tasks that need frame-by-frame precision. The new mode is available now for Gemini 3.8, 3.7, 3.6 Flash and 3.5 Flash-Lite, and developers can turn it on by setting a single processing field in a request. It works with both uploaded video files and public YouTube links, and a single request can mix modes, agentic for a long lecture, static for a short clip attached to the same prompt.

Billing follows the same per-token pricing Gemini already uses, with no separate fee for the feature. The distinction shows up in how tokens are categorized: reasoning about where to look in the video counts as thought tokens, while the frames, audio, or transcript text the model actually pulls in count as tool-use tokens. Developers can confirm agentic mode ran by checking for new processing_call and processing_result entries in the response, which mark when the model requested a segment and when that segment loaded.

Google’s own evaluations put Gemini 3.7 Flash with agentic understanding at the front of the accuracy-to-cost tradeoff among the models it tested, though the company has not published the full benchmark methodology or comparison set. The feature is only available through Google’s hosted API, in AI Studio and the Gemini Enterprise Agent Platform; there are no open weights to run independently. What remains to be seen is how the cost and accuracy figures hold up once developers apply agentic processing to their own video workloads outside Google’s test set.