Quick answer
Google has launched agentic video understanding for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite. Instead of sending frames sampled at a fixed rate into the context in one pass, agentic mode lets the model navigate the timeline, inspect selected video segments, read transcripts, and adjust frame rate and resolution for the question. Google reports token reductions of up to 88%, cost reductions of up to 66%, and accuracy gains of up to 7% in its own evaluations. The feature is available through the Gemini API and Gemini Enterprise Agent Platform, but its availability inside Chat AI has not been verified.
Agentic video understanding is live on three Gemini Flash models
Google says developers can use agentic video understanding today with Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite. The feature works with uploaded video files and public YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It is a video-input analysis feature, not a video-generation model.
Sources: Google, Google AI for Developers
The model chooses which moments and signals to inspect
Gemini's default static mode samples video at a fixed rate of one frame per second and places that material into the context in a single pass. Agentic mode instead runs a goal-directed loop that can search the timeline, load selected segments, inspect audio or transcripts, and change frame rate or resolution as needed. That design is intended to avoid filling the context with frames that are irrelevant to the user's question.
Sources: Google, Google AI for Developers
Google reports lower token use and better accuracy on long-form video
Across Google's standard video-analysis benchmarks, the company reports that agentic processing reduced token use by up to 88%, cut analysis cost by up to 66%, and improved accuracy by up to 7%. Google says the gains are most pronounced on long-form material such as lectures and multi-hour recordings. These are first-party results expressed as maximum improvements, not guarantees for every model, video, prompt, or workload. Teams should measure accuracy, latency, token use, and total cost on their own content.
Sources: Google, Google AI for Developers
Developers enable the mode on each video input
The Gemini Interactions API accepts a processing field on a video input. Setting it to agentic asks the supported model to navigate the media dynamically. Google says developers can verify that the mode ran by inspecting the response steps for processing_call and processing_result entries. Applications can also mix static and agentic processing for different videos in the same request.
Sources: Google AI for Developers
Use agentic mode for long searches and static mode for short, latency-sensitive clips
Google recommends starting with agentic processing when the query targets a specific moment or the video is long enough that loading fixed-rate frames would consume substantial context. Static processing can still be preferable for short clips under five minutes when latency matters or when an application needs consistent frame coverage across the entire clip. Public YouTube input remains a preview feature, and Google's documentation says only public videos can be submitted through that path.
Sources: Google AI for Developers
The API feature is not yet verified inside Chat AI
Google says agentic video understanding uses standard Gemini API token pricing with no extra feature fee. The company plans to roll it out across Flash and Flash-Lite models in the Gemini app and later use it for Ask YouTube. Although Gemini 3.6 Flash appears in Chat AI's current verified model directory, that listing does not establish support for this per-video agentic processing mode. This article therefore does not claim that the feature is available in Chat AI.
Frequently asked questions
What readers usually ask
What is Gemini agentic video understanding?
It is a video-analysis mode that lets supported Gemini models navigate a video's timeline, inspect selected segments, use transcripts, and adjust frame rate or resolution according to the question.
Which Gemini models support agentic video understanding?
Google documents support for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite.
How do developers turn on agentic video processing?
In the Gemini Interactions API, set the video's processing field to agentic. The response steps can be checked for processing_call and processing_result entries to confirm that dynamic navigation occurred.
Does agentic mode always reduce cost by 66%?
No. Google reports reductions of up to 66% in its evaluations. Actual cost and token use depend on the model, video length, content, prompt, processing path, and current API pricing.
Can Gemini analyze YouTube videos in agentic mode?
Yes, Google says the feature supports public YouTube videos through the Gemini API. The YouTube URL input path is in preview and does not accept private or unlisted videos.
Is Gemini agentic video understanding available in Chat AI?
The feature has not been verified in Chat AI. A model listing by itself does not confirm support for Google's per-video agentic processing mode.
Evidence