Gemini's agentic video mode cuts long-video tokens up to 88%
Process hours of footage for a fraction of the token bill, plus open-weights video you can self-host, controllable 4K clips, and a bot that triages your PRs.
Copy markdown
88% fewer tokens, and 7% more accurate
Google's new agentic video understanding lets Gemini actively navigate a clip, deciding what to examine, at what speed, and through which modality (frames, audio, or transcript), instead of ingesting every frame at a fixed rate. Google reports up to 88% fewer tokens, up to 66% lower analysis cost, and up to 7% better accuracy, with the gains concentrated on long videos.
Which models, and how to switch it on
It's live on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via the Gemini API and AI Studio: set video processing to 'agentic' in the request config, at standard token pricing with no extra fee. Think multi-hour search, split-second edit detection, and anomaly spotting without the usual frame-by-frame token blowout.
Elsewhere: LTX-2.5 open-weights video, up to 4K
For the generation half of a video pipeline, Lightricks' LTX-2.5 is on Hugging Face and fal: synchronized video plus audio up to 4K and ~20s on the Fast tier, under a community license that's free below $10M ARR, with API from $0.09/second. Pair it with Gemini's cheap agentic analysis for a full generate-and-check loop this weekend.
Elsewhere: 'PR Sous Chef' pings you only when it matters
GitHub's agentic-workflows team published a reference 'PR Sous Chef' agent that scans open pull requests every 15 minutes, decides which ones actually need a human, and fires a targeted Copilot action only when there's a clear, actionable gap. It's a copyable template for cheap, low-noise agents: read-only triage first, model calls second.
Elsewhere: Gemini Omni 1.1 Flash extends clips to 40s
Google also pushed Gemini Omni 1.1 Flash to general availability: controllable video generation with scene extension out to ~40 seconds, keyframe control, frame interpolation, and 4K output (upscaled). The old gemini-omni-flash-preview endpoint retires September 30, so move production calls to the GA model before then.