Introducing agentic video understanding with Gemini

What it claims

Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite can now process video through an agentic loop instead of fixed-rate frame sampling. The page's own comparison point is "static" processing at a default of 1 FPS (adjustable via the API). Against that baseline it reports up to 88% fewer tokens, up to 66% lower analysis cost and up to 7% higher accuracy "across standard video analysis benchmarks." Google positions Gemini 3.7 Flash with agentic processing as the best-quality option and places it on the accuracy-to-cost Pareto frontier among the models it tested.

How it works

The model calls an internal tool that loads only the parts of the video file it needs. It decides which segment to inspect, at what speed or frame rate, and through which modality: frames, audio or transcript. Google frames this as automating what developers previously did by hand with their own chunking and sampling pipelines.

Two capability descriptions make the mechanism concrete:

• Anomaly detection works by resampling chosen time windows at a frame rate above the default to inspect fast motion or subtle artifacts.
• Moment retrieval targets state changes and cut boundaries that fall between frames at 1 FPS.

How to use it

• Inputs: video uploads and YouTube URLs.
• Platforms: Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
• Pricing: standard Gemini token pricing with no additional feature fee, so the savings come entirely from lower token use.
• Enabling it: set "processing": "agentic" on the video input item. This uses the interactions API rather than generate_content:

from google import genai
client = genai.Client()
interaction = client.interactions.create(
    model="gemini-3.7-flash",
    input=[
        {
            "type": "video",
            "uri": "https://youtu.be/7Z5Vy9JBANs",
            "processing": "agentic"
        },
        {
            "type": "text",
            "text": "What are the 3 most important announcements in this keynote?",
        },
    ],
)
print(interaction.output_text)

The page links to a separate developer guide for further detail.

Limits

• Headline numbers lack context. All figures are "up to." The page does not name the benchmarks, list per-model or per-benchmark results, or state the static FPS used in the comparison. The 88% token and 7% accuracy figures are attributed specifically to 3.7 Flash.
• Long-form bias in the gains. Google says the gains are largest on long videos (10-minute how-tos through multi-hour recordings). The page does not say how short clips fare.
• No latency or variance data. The page gives no figures for latency or for how much token use varies between calls under a dynamic loop. Budgeting from the "up to" numbers is therefore unverified.
• Partner results missing. The "Real-world results" section introduces partner quotes, but none appear in the fetched text.
• Consumer rollout not yet live. Availability in the Gemini app is described as "soon," and Ask YouTube as "in the coming months." Only the API is available now.