How to Choose Gemini Video Mode—Static 1 FPS or Agentic Retrieval

Sending video to the Gemini API is also a choice about how the model turns that file into tokens. The Video understanding docs describe two processing modes. Static is the default for every Gemini model, including Gemini 4: extract frames at about one per second and load them once. Agentic is available on a listed subset of models: the model explores the timeline against the prompt instead of dumping every sampled frame into context up front. Everything below follows Google’s developer docs as of October 26, 2026 (Beijing time). This page covers the Gemini API only—not the Gemini App. For reference only; static vs agentic cost or quality has not been measured on this site. Feedback with production numbers can go through the site-owner contact link in the footer.

Two modes: static vs agentic

Static is simple and predictable. Agentic is a different contract: the model decides which segments and transcripts to load. That can help on long, targeted queries—and it can also add thinking and tool-call tokens when the whole clip still needs to be inspected.

When to pick which

Google’s “Choose a processing mode” guidance lines up with common jobs:

Prefer static when:

• The clip is short (under about 5 minutes).
• A broad full-video summary is needed where most of the recording is relevant.
• The query is latency-sensitive.
• Frame-level precision across the whole clip is required.
• Extra thinking tokens (total_thought_tokens) and tool-call round-trips (total_tool_use_tokens) from timeline navigation should be avoided.

Prefer agentic when:

• The video is long-form (over about 5–10 minutes).
• The task is needle-in-a-haystack retrieval.
• The prompt targets specific moments or topics.

The same docs note that at 1 FPS, fast action or flash-on-screen subtitles can miss detail. If every frame matters on a short clip, stay on static—and raise fps only in static mode when denser sampling is needed.

How to opt in: processing: agentic

Agentic is not automatic. On the video part inside the input array, set "processing": "agentic".

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=[
        {
            "type": "video",
            "uri": video_file.uri,
            "mime_type": video_file.mime_type,
            "processing": "agentic",
        },
        {"type": "text", "text": "What are the three main arguments presented?"},
    ],
    background=True,
)

Modes can be mixed across videos in one request—one part "agentic", another "static". To confirm agentic ran, inspect interaction.steps for processing_call and processing_result.

Static-only knobs: clipping intervals (start_offset / end_offset) and custom fps live under a static processing object. They are not supported in agentic mode.

For long videos or complex prompts where agentic takes more wall time, the docs suggest streaming (stream=True) or background execution (background=True) so the connection does not time out.

Model support is not universal

Agentic is not available on every Gemini model. Per the current docs table, supported rows are:

• Gemini 4 Argon
• Gemini 3.8 Flash
• Gemini 3.6 Flash
• Gemini 3.5 Flash Lite

Static remains the default everywhere else—including Gemini 4 models that are not on that agentic list. Always re-check the current Video understanding table before shipping a mode assumption into production.

Official efficiency claim—for reference only

Google’s table states that agentic can be up to 88% more token-efficient and about ~7% higher quality on long-form content. Those figures are official, self-reported claims. The comparison has not been re-run on this site. Treat them as a vendor reference point when sizing budgets—not as a guarantee that agentic will always be cheaper, faster, or better for a given clip and prompt. On short videos or broad prompts that still require inspecting most of the file, the docs themselves warn that navigation can add thinking and tool-use tokens.

What this is not

• Not the Gemini App. Settings and behavior here are for the Gemini API Video understanding path.
• Not a promise that agentic always wins. Mode choice still depends on length, query shape, and latency budget.
• Not a substitute for Argon plan and release context. Usage ceilings and launch pricing live elsewhere—see the related reads below.

Related reading

Nearby Gemini and harness context on this site:

• Gemini 4 Argon Usage Limits in the Gemini App, Antigravity and API (October 2026)
• Gemini 4 Argon: Release, Pricing, 1M Context
• Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does