Three Gemini Flash Models Add Agentic Video

Gemini 3.7 Flash places the accuracy-to-cost combination at the pareto frontier once agentic video understanding activates.

The launch post states that Gemini 3.7 Flash with agentic understanding offers the best possible quality overall and the best combination of quality and cost efficiency. Static processing ingests video at a fixed frames-per-second rate, defaulting to 1 FPS with manual adjustment through the API. Agentic video understanding instead pairs core reasoning with native video tools that search, scan, and inspect target segments across frames, audio, and transcripts. Rohan Doshi and Mario Lučić wrote in the announcement that the approach cuts token consumption by up to 88 percent, reduces costs by up to 66 percent, and boosts accuracy by up to 7 percent across standard video analysis benchmarks. The same post notes that these gains appear most clearly on long-form content such as 10-minute how-to guides, 90-minute lectures, and multi-hour recordings. The launch post confirms availability today for video uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, with no additional feature fee beyond standard token pricing.

Sub-second moment retrieval becomes possible because the model can resample interesting windows at higher FPS on demand. Long-form needle-in-a-haystack search answers complex queries without ingesting millions of tokens. Anomaly detection and accurate counting of actions or objects follow the same loop. The code example supplied in the developer guide shows the exact configuration required to invoke the feature on a YouTube URI.

Gemini 3.6 Flash receives the identical dynamic scan across frames, audio, and transcripts on its supported tier.

The announcement lists 3.6 Flash among the three models that receive the update. Gains span all three supported models, yet teams already running workloads on the 3.6 tier can obtain the efficiency improvements without shifting to a larger model. The same active tool-calling loop that fetches only needed segments applies here, preserving the reported token reduction of up to 88 percent and cost reduction of up to 66 percent. The launch post records that the feature uses standard Gemini API token pricing with no extra charge. Early access partners tested the capability on long-form material and observed the accuracy lift of up to 7 percent while keeping inference spend lower than static baselines.

Gemini 3.5 Flash-Lite brings agentic processing to the lightest supported Flash variant without an added fee.

The launch post confirms availability on this model alongside the two larger Flash variants. Static processing forces developers to choose between high token costs or techniques that drop critical details on extended recordings. The 3.5 Flash-Lite version applies the same goal-directed inspection of frames, audio, and transcripts at the lowest inference tier among the three. The announcement states that the configuration parameter remains “agentic” and that the capability will roll out to all users in the Gemini app across Flash and Flash-Lite models soon. In the coming months the same processing will power YouTube’s Ask YouTube feature on the video watch page.

Agentic Vision supplies the earlier image counterpart that shares the same Think-Act-Observe loop now extended to video.

The January 2026 post on agentic vision describes how Gemini 3 Flash combines visual reasoning with code execution to zoom, inspect, and annotate images step by step. That earlier capability already delivered a consistent 5-10 percent quality boost across most vision benchmarks once code execution is enabled. The September video announcement builds directly on the same pattern, replacing manual inspection with an internal tool call that loads relevant video segments. Both features therefore rest on the same principle of active investigation rather than single-pass ingestion. DeepMind’s news index places the two posts in the same models category, confirming the continuity of approach.

The image version required explicit enabling of code execution under Tools in Google AI Studio. The video version removes that step by baking the native video tool call into the model itself when processing is set to agentic.

No published implementation yet covers real-time camera streams or live broadcast feeds.

The sources describe support only for uploaded video files and YouTube videos via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The supplied code example uses a static URI and sets processing to agentic. Capabilities such as sub-second moment retrieval and long-form needle-in-a-haystack search therefore apply to stored content rather than continuous input. The field still lacks published numbers for latency or token behavior on live streams.