Gemini video understanding 88% fewer tokens vs Sume Video inspect

The Gemini API changelog for September 1, 2026 says its agentic video understanding approach uses up to 88% fewer tokens for long-form content than static processing. Sume's two read paths do not use model tokens in that way: Video inspect returns a probe, sampled stills and an optional transcript, and Reference ingest returns a timestamped manifest for one clip.

What Google reports

The changelog entry is the only Gemini claim used here: up to 88% fewer tokens on long-form content compared with static processing. This post does not repeat a benchmark or a price, and the figure is an upper bound, not a typical result.

What Sume returns

Both Sume surfaces read one media.sume.com clip that the workspace already owns.

How you control the cost

With Video inspect you choose the program. Omit frames for 8 mid-bin stills, pass frames: false for a probe only, pass { at: [...] } for named times or { fps: n } for a rate no higher than 2. seek: "fast" snaps to keyframes when you are only skimming. A transcript is optional and billed: transcribe: true runs Sume STT 1.0 on the audio.

Probe and stills are billed by their Modal compute rather than by tokens. Reference ingest transcribes only when the track is not silent and speech is detected, and settles that line to zero otherwise.

Which to use

If you want a model to answer an open question about a long video, an agentic video-understanding API fits that job. If you want facts and stills to feed an agent or a remix plan, Sume's surfaces give you structured output you can inspect yourself. They are inputs to a model, not a replacement for one. Compare cost on your own clips, because the two are billed on different units.