Gemini Adds Agentic Video Processing to Cut Token Usage by 88 Percent
Google has launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. The feature scans video segments dynamically, reducing token consumption by up to 88% and API costs by up to 66%.
Impact: High
Why it matters
You can process long-form video archives and sub-second moment queries with a single API configuration change while slashing multimodal token spend.
TL;DR
- 01Setting processing to agentic in the Gemini API cuts long-form video token usage by up to 88%.
- 02Supports Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite without extra feature surcharges.
- 03Enables sub-second state change detection and variable FPS sampling for video search agents.
Key facts
- Token reduction (self-reported)
- Up to 88%
- Cost reduction (self-reported)
- Up to 66%
- Accuracy gain (self-reported)
- Up to +7% on LongVideoBench
- Supported models
- Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash-Lite
Dynamic Inspection vs. Fixed-Rate Sampling
Standard multimodal pipelines ingest video files by extracting frames at a constant 1 FPS. For multi-hour lectures or CCTV streams, this approach rapidly consumes millions of context tokens while remaining blind to sub-second transitions. Gemini's agentic video understanding gives the model access to native inspection tools, enabling it to scan transcripts or audio first, identify target timestamps, and dynamically resample visual frames at variable frame rates only where needed.
Benchmark Numbers and API Activation
Google self-reports the following performance metrics on LongVideoBench and standard video analysis suites:
- Token Reduction: Up to 88% lower token consumption compared to static 1 FPS ingestion.
- Cost Efficiency: Up to 66% reduction in analysis costs under standard token billing.
- Accuracy Gain: Up to 7% improvement in answer precision, particularly for sub-second retrieval, action counting, and anomaly detection.
Developers can enable the mode in the Gemini API by setting the processing configuration parameter to agentic in Google AI Studio or the Gemini Enterprise Agent Platform.
Try it in 2 minutes
interaction = client.interactions.create(
model="gemini-3.7-flash",
processing="agentic",
contents=[video_file, "What are the 3 most important announcements in this keynote?"]
)python
✓ When to use
- Use when searching for specific visual events across multi-hour lectures, security footage, or meetings.
- Use when tracking high-speed motions or anomalies that fall between standard 1 FPS sample intervals.
✕ When NOT to use
- Do not use for ultra-short clips (< 5 seconds) where static 1 FPS ingestion already uses negligible tokens.
- Avoid if your application requires strict deterministic frame-by-frame processing without tool-call latency.
What to do today
- Add processing: 'agentic' to your existing Gemini API client configuration for video inputs.
- Benchmark token consumption on your long-form video pipelines in Google AI Studio to compare costs.
Sources