ALT Flowchart showing how agentic video understanding works with Gemini models. An incoming "Query (video + prompt)" leads into Gemini, which enters a cyclical reasoning loop: Gemini moves to "Think," invokes tools—"get_transcript," "get_frames(start, end, fps)," and "get_audio(start, end)"—and receives an "Observation (video frames, audio, or transcript)" before looping back. Once reasoning is complete, Gemini produces the final "Output (text)."

ALT Scatter plot titled "Agentic video understanding: 3.7 Flash is the frontier on accuracy and cost" comparing AI models on the 1H-VideoQA benchmark. Gemini 3.7 Flash with agentic video understanding ranks highest in the upper right, achieving the best combination of high accuracy (~90%) and lowest cost (~$0.10 per query), outperforming standard Gemini 3.7 Flash, GPT 5.6 models, Claude Opus 5.0, and Grok 4.6.