Google adds agentic video understanding to Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite

Google today released a capability it calls agentic video understanding across three of its latest Gemini models — 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Instead of ingesting video at a fixed frame rate (the default is 1 FPS, adjustable via the API), each model now receives an internal tool that lets it search, scan and examine specific segments across frames, audio and transcript on demand. The stated result: fewer tokens, lower cost and higher accuracy.

In place of passive, fixed-rate media ingestion, the model enters an agentic loop: it decides which portion to load, at what rate, and which modality — visual, audio or text — is relevant at that moment. Developers could previously build this logic themselves, but it is now baked into the model through an internal tool call, removing significant engineering overhead. Google frames the feature as a natural extension of the agentic vision capability it shipped earlier, which combined code execution with native image understanding.

According to Google's benchmarks, Gemini 3.7 Flash with agentic understanding shows up to an 88-percentage-point reduction in token consumption, up to a 66-percentage-point savings in analysis cost, and up to a 7-percentage-point accuracy gain compared with static processing on the same models. The advantage is most pronounced on long-form video — from 10-minute tutorials through 90-minute lectures to multi-hour recordings — where static processing forces developers to choose between high token costs and down-sampling techniques that lose critical detail. Google says Gemini 3.7 Flash sits on the Pareto frontier of accuracy versus cost among the models tested for video understanding.

The capability enables four primary use cases: sub-second moment retrieval, including detection of sharp cuts and state changes that 1 FPS misses, making precise automated editing feasible; needle-in-a-haystack search across hours of footage without consuming millions of tokens; anomaly detection with targeted high-rate resampling of interesting time windows to catch fast motion and subtle visual defects; and action and object counting, tracking repeated motions and distinct objects accurately over time.

The feature is available starting today through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform for all three supported models. Activation is a single configuration flag: `processing: "agentic"`. Pricing remains standard Gemini API token pricing with no additional fee for the feature itself. Google points developers to a technical guide for deeper integration; early-access partners have reported strong performance, though those are first-party accounts rather than independent measurements.