Gemini gives video analysis an agentic mode

Google has added agentic video understanding to Gemini's API, allowing supported models to decide which parts of a video to inspect instead of reading the entire file at a fixed frame rate.

The feature is available for video uploads and public YouTube URLs through Google AI Studio and the Gemini Enterprise Agent Platform. Google says it can cut token consumption by up to 88%, lower analysis cost by up to 66% and raise benchmark accuracy by up to 7%. Those performance figures are Google's benchmark claims. The confirmed product change is a new agentic processing setting for Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite.

From fixed sampling to targeted inspection

Gemini's ordinary video path is static: it extracts frames at one frame per second by default and puts them into the model context in one pass. That can suit a short clip, but it gives an application little room to distinguish a quick visual check from a long search through a lecture, call recording or how-to video.

With agentic processing, Google's developer documentation says the model can navigate a timeline, inspect frames, audio and transcripts, then request more detail around a relevant moment. The approach is designed for tasks such as finding a short event in a long recording, checking fast action at a higher frame rate, counting repeated objects and examining anomalies.

That distinction matters because lower token use is conditional on the question and the video. A request that needs close coverage of an entire short clip may still fit static processing better. Google's developer guide specifically positions static mode for latency-sensitive queries on videos under five minutes or for frame-level precision across the full clip. It recommends agentic mode for long videos and questions about particular moments.

The API exposes a trace, not a hidden shortcut

Developers activate the feature by setting a video's processing field to agentic in an Interactions API request. Google documents two new response-step types:

• processing_call, which records the model asking for a video segment or transcript;
• processing_result, which links the returned material to that request; and
• the ordinary model-output step that follows the gathered evidence.

Those fields give an application a way to show progress or audit which loads occurred. They do not make the model's answer automatically correct. Google describes the steps as a processing trace, while developers still need to test their own prompt, video type and failure cases.

The release also changes the engineering trade-off around longer media. Static processing can force a choice between more frames and a larger context bill. Google says agentic mode calls internal video tools to fetch the segments and signals it considers relevant. That can preserve detail around a short event without forcing every second of a multi-hour recording through the same rate.

Availability is broader than the initial benchmark

Google's launch post names Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The current API guide also lists Gemini 3.8 Flash. The feature works with files uploaded through the Files API, Cloud Storage registration, inline video data and public YouTube URLs, subject to the limits for each route. For example, Google's guide recommends the Files API for large files, long videos and reusable uploads; it lists a 20 GB paid-tier file maximum.

Google says the feature uses standard Gemini API token pricing without an extra feature fee. That does not guarantee a cheaper request in every case: its reported savings come from tested video-analysis benchmarks, and it advises streaming or background execution when complex agentic work takes longer.

The product announcement also says the capability will reach Gemini app users across Flash and Flash-Lite models and later support YouTube's Ask YouTube feature. Those are rollout plans, not evidence that every app user or YouTube viewer has it now.

For teams building video search, review or monitoring tools, the practical question is whether targeted retrieval improves their own accuracy enough to justify a slower, multi-step request. Google's update supplies an API control and observable processing steps; the workload-specific result still needs measurement.