Gemini: Cost-Free Agentic Video Analysis

Key Takeaways

1Google launches agentic video understanding on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite

2Up to 88% fewer tokens and 66% lower costs, with increased accuracy of up to 7% according to Google

3Feature available at no extra cost, activatable via the API for long videos and complex queries

4Use cases: quick search, anomaly detection, counting actions and objects

Why it matters : This advancement enables efficient analysis of long videos while controlling costs and token consumption.

Google offers agentic video understanding across multiple Gemini Flash models. The company announces reductions in token consumption and costs, as well as improved accuracy for analyzing long content. The feature is now available via the API at no additional charge.

Immediate Activation and Unchanged Billing

Agentic video understanding uses the standard token pricing of the Gemini API, with no extra fees. To activate it, simply set the processing to "agentic" in the API configuration. The feature is available in Google AI Studio and the Gemini Enterprise Agent Platform, and applies to uploaded videos as well as YouTube content via the API.

From Static to Agentic: The Model Chooses What to Watch

Static processing ingests video streams at a fixed rate, defaulting to 1 FPS, adjustable via the API. The agentic approach combines the model's reasoning with native video tools to dynamically search and inspect images, audio, and transcriptions. The model decides what to analyze, at what speed, and through which channel, retrieving only the necessary moments. It can invoke an internal tool to load the relevant part of the file in an agentic loop, which, according to Google, reduces development burdens compared to manual orchestration.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Claimed Gains and Efficiency on Long Videos

Google reports reductions in token usage of up to 88% and analysis costs of up to 66% on video analysis benchmarks, with accuracy improvements of up to 7%. These gains are particularly notable on long content, from approximately 10-minute guides to 90-minute lectures and multi-hour recordings. The company presents the agentic approach as an alternative to static processing, which often forces a choice between high token costs or methods that omit important details.

Announced Use Cases: Quick Search, Anomalies, and Counting

Highlighted capabilities include retrieving moments in under a second, more precise anomaly detection, and accurate counting of actions and objects. Google cites the identification of state changes in a fraction of a second and difficult-to-see cuts at 1 FPS, responding to complex queries in multi-hour videos without consuming millions of tokens, and resampling time windows at a higher rate to inspect rapid movements and subtle artifacts. Partners with early access report good performance during testing. The feature is launched on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite and is activated via the "agentic" mode of the API.