Google Adds Agentic Video Understanding to Gemini, Cutting Token Use by Up to 88%

Google DeepMind launched agentic video understanding for Gemini on September 1, 2026, a processing mode that lets models actively decide which parts of a video to inspect rather than ingesting footage at a fixed frame rate. The feature is available now for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

What changed

Standard Gemini video processing samples frames at a default rate of 1 frame per second, adjustable via API. That static approach forces a tradeoff: developers either pay for dense frame sampling across an entire video or risk missing details by sampling sparsely. Agentic video understanding replaces this with a dynamic tool-calling loop — the model invokes an internal tool to load specific video segments, adjusts frame rate on the fly, and pulls from visual frames, audio, or transcripts depending on what a query requires.

According to Google, this reduces token consumption by up to 88% and analysis costs by up to 66% on standard video benchmarks, while improving accuracy by up to 7%. The company says gains are most pronounced on long-form content — 10-minute how-to videos, 90-minute lectures, and multi-hour recordings — where static sampling previously forced a choice between high token costs and dropped detail.

On LongVideoBench, a long-form video understanding benchmark, Google reports Gemini 3.7 Flash with agentic processing shows large token reductions alongside accuracy gains compared to static processing, though the company did not publish exact before-and-after benchmark scores in the announcement. Google states that Gemini 3.7 Flash with agentic understanding sits at the accuracy-to-cost pareto frontier among the tested models.

New capabilities

Google lists four use cases enabled by the feature:

• Sub-second moment retrieval: identifying split-second state changes and cut boundaries missed at 1 FPS sampling.
• Long-form needle-in-a-haystack search: answering queries across multi-hour videos without consuming millions of tokens.
• Anomaly detection: resampling specific time windows at higher frame rates to catch rapid motion or visual artifacts.
• Counting actions and objects: tracking repeated movements or distinct objects over time by rewatching segments at variable frame rates.

Pricing and access

Agentic video understanding uses standard Gemini API token pricing with no additional feature fee, according to Google. Developers enable it by setting the "processing" parameter to "agentic" in API calls, as shown in Google's sample code using the genai Python client with model="gemini-3.7-flash".

Google says the feature will roll out to all users in the Gemini app across Flash and Flash-Lite models "soon," and will power YouTube's "Ask YouTube" feature on video watch pages in the coming months.

What this means

This is a processing-mode change layered onto three existing Gemini model checkpoints, not a new model release — the underlying weights for 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite are unchanged. The significance is economic: video understanding has historically been token-expensive because every frame consumes context regardless of relevance to a query. If Google's claimed 88% token reduction holds up under independent testing, it materially lowers the cost of building products that reason over long video — security footage review, lecture search, video editing tools — without waiting for cheaper base models. The zero-fee pricing structure also means Google is betting on volume: cheaper per-query costs should drive more video API traffic rather than compress margins. The open question is how these efficiency numbers translate outside Google's own benchmark selection, since no third-party verification exists yet.