Alphabet’s Google introduces agentic video understanding in Gemini, aiming to cut video-analysis costs sharply

Google is rolling out a new capability for its Gemini family of models that changes how video is analyzed. Called agentic video understanding, the feature gives Gemini a more active role in reviewing video by dynamically scanning relevant segments across visual frames, audio, and transcripts, rather than processing a video at a single preset sampling rate.

In a post describing the launch, Google says the new approach reduces token consumption by up to 88% and lowers costs by up to 66%, while improving output quality by up to 7% on video analysis benchmarks. The company frames the update as both an accuracy enhancement and a cost-control mechanism, especially for long-form content where fixed-rate “static” processing forces teams to trade off between spending and completeness.

Agentic video understanding is being introduced for three Gemini variants: Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite. Google also says Gemini 3.7 Flash with agentic understanding delivers the best overall quality and the strongest mix of quality and cost efficiency among the tested models for video understanding, positioning it at the “accuracy-to-cost pareto frontier” according to the company’s benchmark comparisons.

The company draws a contrast with what it calls static processing. In the static approach, Google says the model ingests video media streams at a fixed frames-per-second rate, defaulting to 1 FPS and adjustable via the API. With agentic video understanding, Google says Gemini pairs its core reasoning with native video tools to search, scan, and inspect targeted segments, fetching only the moments and indicates it needs.

Google adds that the capability is designed to work on both visual and non-visual elements of video. By using modalities including frames, audio, and transcripts, the system can support tasks such as sub-second moment retrieval, more accurate anomaly detection, precise counting, and other video-processing workflows that depend on finding specific events within a clip. The point, according to Google, is to reduce wasted processing on irrelevant parts of a video while improving how precisely the model homes in on what developers ask for.

A practical implication of the change is how developers build for long-form video. Google says the efficiency gains are most pronounced as videos get longer, citing use cases that span 10-minute tutorials through 90-minute lectures and multi-hour recordings. For these workflows, the company argues that static processing can become expensive or can drop important details, while an agentic approach can be more selective about what it examines.

Google says the feature is available now via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. To enable it, developers set processing to “agentic” in the API configuration. The company also states that it uses standard Gemini API token pricing and charges no additional feature fee for agentic video understanding.

The release also connects the update to broader product experiences. Google says it will roll out agentic video understanding to all users in the Gemini app across Flash and Flash-Lite models “soon.” Separately, Google says that in coming months the capability will power YouTube’s “Ask YouTube” feature on the video watch page, using Gemini to produce higher-quality answers grounded in the visuals.

Google did not provide detailed information in its announcement about the exact evaluation methodology behind the benchmark claims, nor did it disclose the specific workloads or datasets used to estimate token savings and cost reductions. While it states the results hold across standard video analysis benchmarks and names the approximate ranges of improvement, developers may still need to test the feature on their own content types and production constraints to confirm performance and cost impacts.

As Google’s video-understanding push moves from research to deployment, the next question for customers will be how the agentic approach behaves across different video characteristics, including speech quality, transcript availability, and event density. For API users, the immediate watch item is how quickly agentic processing becomes a default option in more Gemini configurations and how the new capability affects latency and developer effort in real-world pipelines. For YouTube, the key will be whether “Ask YouTube” can reliably ground answers in the relevant moments without inflating compute costs for longer uploads.