It says the model can search timestamped transcripts, fetch targeted frame sequences at adaptive frame rates, and extract audio through a server-side tool loop. The guide lists Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite as supported models, with the API option processing="agentic".
The thesis: video analysis can shift from bulk ingestion to query-directed retrieval. The mechanism: a lightweight video reference is passed first; the model plans what evidence it needs, retrieves selected modalities or time windows, observes those results, and can repeat the cycle before answering. Why now: Google is exposing this loop as an API mode for its Flash models, so developers can test it in applications rather than treat it only as an internal technique. The first implication: teams handling long recordings can compare selective retrieval against static sampling on their own workloads. Teams handling long recordings can compare selective retrieval against static sampling on their own workloads.
Google reports up to 88% fewer tokens, up to 66% lower analysis cost, and up to 7% higher accuracy on fine-grained visual reasoning tasks versus static 1 FPS processing. These are vendor-reported benchmark maxima; the guide does not establish that every video, query, or production workload will see those gains. The evidence boundary: the supplied source is Google’s developer guide, not an independent replication or a customer deployment report. The first watch condition: measure end-to-end accuracy, latency, and total cost against static processing on representative long-form videos before making the mode a default.