Top AI Product

Video is the most expensive thing you can feed an LLM. A single query over a long video used to eat 300–400K tokens because models brute-forced every frame at a fixed 1 FPS. Google’s answer, launched September 1: stop telling the model how to watch. Gemini Agentic Video hands that decision to the model itself — skim or go frame-by-frame, watch the pixels or listen to the audio track, zoom into a suspicious window and resample it at higher frame rate.

What it actually does

It’s an API capability, not a new model. On long-video benchmarks like 1H-VideoQA and LVBench, per-query tokens drop from ~350K to under 50K — 88% fewer tokens, 66% lower cost — and accuracy goes up 7%. Cheaper and smarter at once is rare. It handles sub-second moment retrieval across multi-hour footage, anomaly detection, and counting repeated movements.

How to use the API

Available now in the Gemini API on 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite: set processing to agentic in your config. Standard token pricing, no feature fee. Typical use cases: security-footage review agents, sports analysis, searching hours of screen recordings.

You Might Also Like

• Google Releases Gemini Embedding 2 one Vector Space for Text Images Video and Audio
• Google Gemini Omni Flash Drops Video Generation to 0 10 a Second
• Google Gemini 3 5 pro ga Ships With a 2m Token Context Window the Biggest in any Production Model
• Google Gemini Enterprise Gives Every ai Agent a Cryptographic id its Real Fight With Openai and Anthropic
• Google Gemini Spark the Gemini app Just Became a 24 7 Agent That Runs Without Your Phone