共有:

Google DeepMind / Gemini Flash

Video goes from "watch it all"
to "watch what matters"

Feeding a long video straight into a model used to mean paying for every frame in input tokens. Google DeepMind just answered that problem head-on: the new video understanding in Gemini Flash stops blindly ingesting frames and lets the model itself decide what to look at.

AI Navigate Editorial2026.09.066 min read

Before: fixed 1fps reads the whole clip evenly Frames read scale with video length switch to agentic mode After: only the relevant segments are read Frames read scale with how much of the clip matters to the query
01
Why It Matters Now

Video was always the most
token-hungry input

Unlike text or images, video grows heavier the longer it runs — and that weight has always driven the API bill.

Until now, Gemini ingested video at a fixed rate (one frame per second by default), sampling on a metronome regardless of what was actually in the footage. It was predictable, but on anything longer than a couple of minutes, most of those 1fps frames turned out to be redundant or irrelevant to the question being asked — a structural waste. The change detailed on Google's official blog tears up that "sample everything evenly" assumption entirely.

In the new approach, the model scans across visual frames, audio, and transcripts, using native video tools to zoom, scrub forward and back, and resample — pulling in only the segments relevant to the query. This isn't a minor tuning tweak; it's a shift in the fundamental ingestion strategy, from "always sample everything" to "the model chooses what to look at" — the same agentic turn that's already reshaped search and browsing tools now reaching video.


Not watching everything, always —
watching only what matters.


02
By The Numbers

The reductions, by the numbers

On September 1, 2026, Google DeepMind rolled out "agentic video understanding" to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite.

up to 88%
fewer video-input tokens
up to 66%
lower API cost
up to 7%
higher accuracy on standard benchmarks

Up to 88% lower token consumption, up to 66% lower API cost, and up to 7% higher accuracy on standard video benchmarks — those are the figures reported by MarkTechPost's write-up, and every one of them is a ceiling, not a flat guarantee: the gains are largest specifically on long-form video. A thirty-second clip may not see anywhere near that much of a cut.

03
How It Works

The model chooses what to watch

It scrubs, zooms, and resamples across visual frames, audio, and transcripts to narrow in on the relevant segments.

video frames audio transcript AGENTIC SCAN zooms into segments of interest scrubs forward and back resamples frames cross-checks audio and captions Output: relevant segments only the rest is never ingested
FIG. Visual frames, audio, and transcripts are scanned together, and only the segments relevant to the query pass through to the output

The feature applies to three models — Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite — and is live through both the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It's enabled with a single configuration change — setting the processing mode to "agentic" in the API call — with no additional fee for using the feature itself. That's a one-line change to an existing call, which lowers the bar to trying it considerably.

Readers tracking model versions should note one nuance: the source material for this launch names only 3.7/3.6/3.5 Flash-Lite explicitly. The newly-launched 3.8 Flash isn't named in the primary announcement, so support may differ by model ID even within the same "Flash" family — worth checking directly against whatever model ID you're actually running.

Models named in the announcementWhat to verify
Gemini 3.7 FlashSupported (toggle via processing mode)
Gemini 3.6 FlashSupported (toggle via processing mode)
Gemini 3.5 Flash-LiteSupported (toggle via processing mode)
Gemini 3.8 FlashNot named in primary source — verify separately
04
Who It Helps

Who benefits, and how

The upside goes to whoever processes a lot of video. Text-heavy workloads barely notice.

Engineers running video pipelines

Content moderation, video QA, media indexing, and surveillance-footage review all lean on long-form video at volume — exactly where the gains are biggest. Flipping one configuration flag can cut API cost without touching accuracy.

Business stakeholders watching the bill

Anyone funding a large video-analysis workload should see that specific API line item drop. If your product is mostly text-based, this release is essentially irrelevant to your bill — no need to rush a response.

05
What To Do Next

Recommended first steps

01

Benchmark accuracy and cost on your own footage

"Up to 88%" and "up to 7%" are ceiling figures from Google's own benchmarks, not a guarantee for every clip. Before flipping production traffic, run a side-by-side test — support-call recordings, surveillance footage, whatever you actually process — comparing accuracy and cost with processing mode set to agentic.

02

Change exactly one setting

Enable it via the Gemini API in Google AI Studio or the Gemini Enterprise Agent Platform by setting processing mode to "agentic" in the API call. There's no extra fee, so testing it in a non-production environment is low-risk.

03

Choose which workloads to switch

Roll it out first to workloads where the accuracy trade-off checks out. Set aside — and evaluate separately — any workload where a requirement is proving that every frame was mechanically processed.


06
Counterpoint

The risk in "choosing what to read"

The cost savings come with a trade-off that's easy to gloss over. Agentic video understanding narrows what it reads by having the model itself judge what's relevant to the query. When that judgment is right, efficiency goes up — but when it's wrong, the structural risk of skipping a segment that actually mattered doesn't go away.

That matters most for compliance or forensic use cases, where the requirement is proving that the entire video was processed without gaps. For those workloads, a system that lets the model decide what to skip could carry more risk of failing the requirement than it saves in cost — and paying more for exhaustive, non-selective processing may simply be the correct call there.

It's also worth noting that every published figure here is a "ceiling" number from Google's own internal benchmarks; independent third-party verification isn't widely available yet. How well these numbers hold up in production is best judged once outside reviews and real user reports start to accumulate.