Google Adds Agent-Based Video Analysis to Gemini Flash, Cutting Token Usage by Up to 88 Percent

Google is equipping its Gemini Flash models with agent-based video analysis that lets the model decide which parts of a video to examine, rather than processing footage at a fixed frame rate. According to Google, the approach cuts token usage by up to 88 percent, reduces costs by 66 percent, and slightly improves accuracy on internal benchmarks.

The capability is live now for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Developers enable it by setting the processing mode to "agentic" in the API config. Google says there is no additional fee — the feature runs at standard Gemini API token rates.

How static processing worked before

Since native video analysis launched in 2025, Gemini has defaulted to sampling video at one frame per second, transcribing audio and analyzing frames on that fixed schedule. The rate was adjustable through the API, but developers had to choose between high token costs for dense sampling or losing detail with sparse sampling. Fast events — cuts, state changes, or moments shorter than one second — could slip through undetected at the standard 1 fps rate.

What changes with agent-based analysis

Instead of loading an entire video at a fixed rate, Gemini now uses an internal retrieval loop to pull only the frames, audio, or transcript segments relevant to a given task. The model decides autonomously which sections to inspect, at what speed, and through which modality. For suspected anomalies, it can resample specific time windows at a higher frame rate rather than reprocessing the full video.

Google says this lets the model:

• Detect moments shorter than one second, including cuts and state changes missed by 1 fps sampling
• Locate specific scenes across hours of footage without processing the entire file
• Count repeated movements and track individual objects over time
• Support more precise automated video editing

The feature builds on "agentic vision," which Google shipped for Gemini 3 Flash in January 2026. That earlier release let the model write and run Python code to zoom, crop, and annotate still images in a think-act-observe loop. Google had flagged extending this reasoning approach to video as a planned capability when it announced Gemini 3 Flash in December.

Benchmark claims

On Google's internal 1H-VideoQA and LVBench evaluations, the company reports an 88 percent drop in token usage alongside a slight accuracy increase. Google also states that Gemini 3.7 Flash with agentic processing achieves the highest overall quality score on LongVideoBench and the best accuracy-to-cost ratio on 1H-VideoQA among the models tested. These figures come from Google's own benchmarking and have not been independently verified.

Rollout plans

The agentic mode is available now via the API for developers working with uploaded video files and YouTube links. Google says the feature will reach all Gemini app users on Flash and Flash Lite "soon," and will power the "Ask YouTube" feature on YouTube's playback page "over the coming months," tying answers more directly to visible video content rather than transcripts alone.

What this means

Long-form video has been one of the most token-expensive input types for multimodal models, since fixed-rate frame sampling scales linearly with video length regardless of how much actually matters in a given clip. If Google's 88 percent token reduction holds up under independent testing, it substantially lowers the cost of building products around hour-plus video — lecture search, surveillance review, sports analytics, editing tools — without forcing developers to sacrifice detail on short, fast events. The bigger signal is architectural: Google is pushing more of its Flash line toward agentic, tool-using inference loops rather than static context stuffing, a pattern likely to spread to text and image processing as well.