BitShovel
中

alphasignal-daily snapshot and original · Published 2026-09-02

Gemini video understanding: find the relevant moments, process less

Gemini can choose video segments, frames, audio and transcripts in response to a question. Processing less material can reduce long-video costs; standard token prices have not been cut by this feature.

Let the question guide viewing
BitShovel diagram · based on the article’s sources. · Open full-size image

中文

Find moments that answer the question

For a lecture question, a useful approach is to locate the relevant transcript passage and then inspect its frames. Fast movement may need denser sampling in a short window. These illustrate the workflow, rather than a BitShovel test. Static mode samples at a fixed rate, defaulting to one frame per second.

The September 1 launch covered Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The documentation checked for this update also lists 3.8 Flash. Set processing to agentic on the video input. Uploads and YouTube links are supported; the announcement described Gemini App and Ask YouTube integration as future rollouts.

中文

Match the mode to the video

Google reports up to 88% fewer tokens, 66% lower analysis cost and 7% higher accuracy. These maxima are not a combined guarantee for every video. Standard API token pricing applies without an extra feature fee; the original benchmark chart remains below.

The guide favors agentic mode for long videos and targeted moments. Static mode can suit short, latency-sensitive clips or full-clip frame coverage. Navigation adds reasoning and time. The processing_call and processing_result response steps confirm dynamic loading; verify the answer against the video itself.

Google benchmark comparison of token usage and accuracy for agentic video understanding
Google’s September 1 chart compares token use and accuracy for Gemini 3.7 Flash across three video benchmarks. Conditions and results are vendor-reported, not measurements for arbitrary videos. · Open full-size image

Important additions and corrections

Dates below mark changes to our coverage.

  • Correction ·

    Separated processing cost from token rates and added mode choice, dynamic loading and current documentation scope.

Sources and notes

sources中文

Sources and notes

Retained summary and context

Cost and accuracy improvements are vendor benchmark maxima, not BitShovel measurements. Standard token rates are unchanged; future product integrations are separate from available APIs.

Sources and evidence · 3 links

Each link has a purpose and scope. Link counts do not establish independent confirmation; direct support applies only to the statements specified below.

Source timeline
  1. · Found this reading link via AlphaSignal daily.
  2. · Corroboration added: Google DeepMind blog carried the same story this window (title match).
Collection record

First discovered:

Collection and source publication have distinct dates.

Editorial record

Corrected

· Beijing time (UTC+8)

Separated processing cost from token rates and added mode choice, dynamic loading and current documentation scope. Event, discovery and collection dates remain unchanged.

Sources and full record

This dates our review of the wording and sources, not a project release or product update.

Full record: before, after and sources (JSON)

Corrections and feedback