Gemini API adds agentic video processing with up to 88% fewer tokens

• Video analysis can now choose frames, audio, and transcript segments around the question being asked, Google DeepMind's rollout thread says, rather than applying one uniform scan rate.
• Google measured up to 88% lower token use and up to 7% higher quality for Gemini 3.7 Flash: OfficialLoganK's benchmark card reports the within-model comparison.
• API callers set processing: "agentic" on an individual video, and Google DeepMind's rollout thread lists Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite as the supported models.
• Video uploads and public YouTube URLs are available through the API today, Google DeepMind's rollout thread says, with no separate feature fee under Google's standard token pricing.

The developer guide puts a lightweight video pointer and metadata in the initial context, then retrieves timeline segments on demand. Its mixed-mode example runs a 45-minute reference match agentically and a 10-second query clip statically. bilawalsidhu asked Google for an open-source video-editor reference implementation in a reply.

Google's developer guide calls the server-side process a Think, Act, Observe loop. Frame selection is model-directed, OfficialLoganK's reply said.

1. Pass by reference: the initial context contains the video ID, duration, and container metadata.
2. Plan: Gemini uses the prompt to decide which timestamps and modalities can answer it.
3. Fetch slices: it can search timestamped transcript text, pull audio, or load selected frame sequences. The guide gives 0.1 FPS for skimming and 5 to 10 FPS for fast motion as examples.
4. Observe and repeat: retrieved material becomes new context, allowing another targeted search before the final answer.

Google's launch post compares 3.7 Flash with its agentic setting against static processing at high thinking, low media resolution, and 1 FPS.

• Minerva: 80.9K to 33.6K tokens, 58.4% fewer; accuracy 73.7% to 79.0%, +5.3 points.
• 1H-VideoQA: 397.6K to 47.7K tokens, 88.0% fewer; accuracy 87.5% to 88.5%, +1.0 point.
• LVBench: 300.3K to 36.0K tokens, 88.0% fewer; accuracy 85.1% to 88.6%, +3.5 points.

The processing field belongs to a single video input, as OfficialLoganK's launch post notes, so one request can mix modes across assets.

Google's Cloud documentation explicitly supports an agentic long recording beside a static short clip in the same request. The developer guide says billing covers the media tokens the model actually materializes, plus standard thinking and output tokens; Batch API requests run at half the standard cost.

Google's launch post identifies four video-analysis jobs for the mode:

• Find split-second state changes and tight cut boundaries.
• Search multi-hour recordings for a specific event or answer.
• Re-sample an interesting window at higher FPS to inspect rapid motion or subtle artifacts.
• Track repeated actions and distinct objects over time.

gregisenberg described that range of source material, from meetings and podcasts to product demos, training footage, security video, and sports, as a new kind of database in a response.

Google's launch post says agentic video understanding will reach Gemini app users on Flash and Flash-Lite soon. It also names YouTube's Ask YouTube, on the watch page, as a forthcoming use of the feature for answers grounded in a video's visuals.