Gemini vs Hugging Face

Agentic video understanding is now available on Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, switched on with an API config setting. Instead of ingesting video at a fixed frame rate, the model uses native video tools in an agentic loop to search, scan and inspect segments across frames, audio and transcript, cutting token consumption by up to 88 percent and cost by up to 66 percent while improving accuracy by up to 7 percent. It follows a fortnight of speech work — a dedicated 3.5 Transcribe model, voice-driven task delegation in Gemini Live — and a developer model refresh in Omni 1.1 Flash. Feed bodies are teaser length; this window was read from the source posts.

The same pattern keeps repeating at model level: give the model a native tool and a loop, and let it decide how to use its own context. Computer use in June, robotics in July, agentic vision for images, and now video. Google names agentic vision as the direct precedent for this release, so the technique is established and video is the modality it just reached — which is also why the efficiency numbers, not the capability, carry the announcement. The application layer is running a separate clock, converging on voice as the way input arrives across Live, Workspace and macOS.

Agentic vision covered images and this covers video, so audio-only and document processing are the modalities the same loop has not yet been pointed at. Whether the setting becomes the default rather than an opt-in config value is not stated.