Gemini 3.7 Flash decides which parts of a video to watch, cutting tokens up to 88%

Google launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite (Google). Instead of scanning a video start to finish at a fixed sample rate, the model runs an internal loop deciding what to watch, at what speed, and through which channel: frames, audio, or transcript. It fetches only the segments needed to answer the query.

Google reports up to 88% lower token consumption, up to 7% higher accuracy, and about 66% lower cost, placing it on the accuracy-to-cost pareto frontier for the video models they tested. Live now in the Gemini API through AI Studio and the Gemini Enterprise Agent Platform, consumer app to follow.

Accuracy going up while tokens go down is the part that made me stop. Every compression technique I've shipped traded one for the other. This one doesn't, and the reason is that fixed-rate scanning was never a good use of the budget in the first place. If the answer to your question lives in eleven seconds of a forty-minute video, sampling the other thirty-nine minutes at 1fps is not context, it's noise you're paying to carry.

The mechanism is the same selective-retrieval move showing up everywhere in agent context work right now. LatentPress writes conversation histories into continuous memory tokens fed straight through the decoder's input-embedding interface, hitting 0.504 accuracy on LongMemEval at 7.70x compression against 0.490 for uncompressed evidence (arXiv 2609.01507). ACToR fires retrieval at the specific decode positions where repository-level code generation goes wrong, for 15.4% relative gain on CoderEval over task-level RAG (arXiv 2609.01601). Different domains, same insight: deciding what to look at beats looking at everything, and it's cheaper.

For anyone doing video at volume, run the numbers this week. A 66% cost cut on a workload that was previously too expensive to run at all changes which products are buildable. I've had two ideas sitting in a notes file for a year that died on video token cost, and I'm going back to look at them.

The caveat: these are Google's own numbers on Google's own tested set. "Up to 88%" is a ceiling, not an average, and it will depend entirely on how localized the answer is in your videos. A query that needs the whole video still needs the whole video.