Gemini Flash models get agentic video understanding
Google DeepMind has enabled agentic video understanding on its Gemini Flash models, letting them selectively fetch video segments rather than ingest frames at a fixed rate — cutting reported token use by up to 88%.
Google DeepMind said Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite can now process video "agentically," replacing fixed-rate frame ingestion — the default is 1 FPS, adjustable via API — with a loop in which the model decides what to watch, at what speed, and through which modality (frames, audio, or transcript), invoking an internal tool to load only the needed segments.
Across unnamed "standard video analysis benchmarks," Google reports up to 66% lower analysis cost, up to 88% fewer tokens, and up to 7% higher accuracy. It says the gains are largest on long-form content — 10-minute how-tos to 90-minute lectures and multi-hour recordings — where static processing forces a choice between token cost and dropped detail. Google claims Gemini 3.7 Flash with the feature sits at the accuracy-to-cost Pareto frontier among tested models.
Cited capabilities include sub-second moment retrieval, needle-in-a-haystack search over multi-hour video, anomaly detection via higher-FPS resampling of selected windows, and object and action counting.
The feature is live for uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, enabled by setting processing to "agentic". It uses standard Gemini API token pricing with no additional fee. Google says it will reach Gemini app users on Flash and Flash-Lite "soon," and will power YouTube's "Ask YouTube" watch-page feature in the coming months.
Google names no specific benchmarks, baselines, or evaluation methodology, and its post references early-access partner results without including any partner statements.