01
What is Gemini agentic video understanding?
Google launched agentic video understanding on September 1, 2026 for Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Instead of processing a video only through the normal fixed sampling path, supported models can dynamically navigate the timeline and request visual frames, audio or transcript context based on the prompt.
Static video processing remains the default. Google's current guide describes static mode as extracting frames at a fixed rate of 1 FPS and placing the result into context in one pass, while agentic mode can adapt frame rate and resolution as it searches for relevant moments.
02
What the agentic mode is designed to do
The new mode is aimed at long-form video and targeted retrieval. Google highlights sub-second moment retrieval, anomaly detection, precise counting and questions where a model needs to revisit only a narrow part of a long recording.
For creators and marketers, that can support workflows such as finding every mention of a topic in a podcast, locating a concise explanation inside a keynote, identifying repeated themes across a long interview or surfacing candidate timestamps for a separate editing workflow.
03
Token efficiency and benchmark claims
Google reports up to 88% lower token consumption, up to 66% lower analysis cost and up to 7% higher quality on evaluated video-analysis workloads. These are Google-reported benchmark results, not independent guarantees.
Real token use remains variable because agentic mode loads content based on query complexity and dynamic sampling depth. A prompt that requires detailed inspection across much of a video can still consume substantial context.
04
Gemini agentic video pricing
Google says agentic video understanding uses normal Gemini API token pricing and has no separate feature fee. For Gemini 3.7 Flash, the current standard paid rate is $0.75 per million input tokens and $3.75 per million output tokens.
Do not turn Google's benchmark cost reduction into a fixed discount assumption. The actual cost depends on the selected model, prompt, video, input method and how aggressively the model resamples parts of the timeline.
05
Video input options and limits
Google supports video through the Files API, Cloud Storage registration, inline data and public YouTube URLs. The video guide currently lists Files API limits of 20 GB on paid access and 2 GB on free access, and recommends the Files API for large, long or reusable videos.
YouTube URL input is currently labeled Preview and Google says its pricing and rate limits are likely to change. Only public YouTube videos are supported by that path.
06
How developers enable it
In the Interactions API, the key change is setting the individual video's processing field to agentic. The processing mode is set per video, so a single request can mix an agentic long recording with a static short clip.
Google documents Python, JavaScript and REST examples. The dedicated implementation guide below covers the setup in more detail.
07
Availability and current rollout
The feature is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Google says consumer Gemini app rollout across Flash and Flash-Lite models is coming soon, so that rollout should not be described as universally complete yet.
Google also says the capability will later power YouTube's Ask YouTube experience. That is a forward-looking rollout statement, not evidence that every YouTube surface is already using the feature.
08
Important limitations
Agentic video understanding analyzes video; it does not generate or edit final video output. A separate editing or generation system is still required for production media workflows.
Static mode can still be appropriate for short clips or tasks where predictable broad coverage matters more than selective retrieval. Production teams should compare task accuracy, latency and token use on their own representative videos.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.