Google Gemini 推出智能体式视频分析，token 用量最多降 88%

Aioga Editorial Team

AI资讯 / 行业动态

返回 AI资讯

Google Gemini 推出智能体式视频分析，token 用量最多降 88%

The Decoder：AI News（RSS）Aioga 编辑团队2026-09-02T08:21:38.000Z热度 58

Google 为 Gemini Flash 系列模型（Gemini 3.7 Flash、3.6 Flash 和 3.5 Flash-Lite）推出智能体式视频分析，模型自主决定...

行业动态The Decoder：AI News（RSS）

今日 AI 情报摘要

Google 为 Gemini Flash 系列模型（Gemini 3.7 Flash、3.6 Flash 和 3.5

Flash-Lite）推出智能体式视频分析，模型自主决定查看哪些片段、以何种帧率和模态获取信息，取代原来每秒固定采样一帧的静态处理。

中文正文 · AI 翻译

谷歌正在将基于代理的视频分析功能添加到多个Gemini模型中。与以固定速率逐帧扫描视频不同，该模型会自行搜索相关片段，谷歌表示，这大大减少了令牌使用量和成本。

最新的模型 Gemini 3.7 Flash：https://the-decoder.com/gemini-3-7-flash-lands-with-coding-gains-and-undercuts-its-three-week-old-predecessors-price-by-50/、3.6 Flash 和 3.5 Flash-Lite：https://the-decoder.com/googles-gemini-3-5-flash-follows-anthropic-and-openai-in-making-newer-ai-models-significantly-pricier/，能够捕捉短于一秒的瞬间，包括以每秒一帧的方式会漏掉的状态变化或镜头切换。Google 表示，这使自动化视频编辑的精度大幅提高。

该系统还可以在数小时的录像中找到各个场景，而无需消耗数百万个令牌。它通过以更高帧率重新采样可疑时间窗口来发现异常，并准确统计重复动作和随时间变化的各个对象。广告

Zwei Balkendiagramme vergleichen Gemini 3.7 Flash mit und ohne agentische Verarbeitung auf Minerva, 1H-VideoQA und LVBench, links Tokens pro Anfrage, rechts Genauigkeit.

直到现在，Gemini依赖静态处理，默认以每秒一帧的固定帧率对视频进行采样，并可通过API进行调整。自2025年本地视频分析推出以来：https://the-decoder.com/googles-gemini-models-add-native-video-understanding/，Gemini一直在对音轨进行转录，并按每秒分析视频帧。广告

谷歌表示，基于代理的变体将模型的推理直接与本地视频工具关联起来。模型自行决定查看哪些片段、以何种速度以及通过哪种模式，无论是帧、音频还是转录。它只提取完成特定任务实际需要的瞬间和信号。

Gemini现在使用内部工具仅抓取视频文件的相关部分。开发人员以前也可以手动构建这种选择性方法，但现在模型可以自行处理。广告

Ablaufdiagramm einer Gemini-Schleife aus Query, Think, Observation und Output mit den Werkzeugen get_transcript, get_frames(start, end, fps) und get_audio(start, end).

该方法基于“能动视觉”（agentic vision），Google 已于一月为 Gemini 3 Flash 推出了这一功能：https://the-decoder.com/google-deepmind-gives-gemini-3-flash-the-ability-to-actively-explore-images-through-code/。该功能允许模型编写并运行 Python 代码来缩放、裁剪和标注图像，在响应前通过“思考-行动-观察”循环检查每个结果。发布时并非在每种情况下都能自动工作，但基础已经建立。当 Google 今年十二月宣布 Gemini 3 Flash 时：https://the-decoder.com/google-makes-gemini-3-flash-the-default-for-search-and-slashes-reasoning-costs/，公司标注了视频的视觉和空间推理作为即将推出的功能。

效率提升在长视频中最为明显，无论是 10 分钟的教程，还是 90 分钟的讲座及多小时的录制内容。使用静态处理时，开发者不得不在高令牌成本与丢失重要细节的方法之间做出选择。在 Google 自家的基准测试中，包括 LongVideoBench，搭载基于代理分析的 Gemini 3.7 Flash 在整体质量评分最高，并提供了准确性与成本效率的最佳组合。广告

Streudiagramm mit Kosten pro Anfrage auf der X-Achse und Genauigkeit auf der Y-Achse; Gemini 3.7 Flash mit agentischer Verarbeitung liegt bei rund 90 Prozent über GPT 5.6 Terra, Claude Opus 5.0 und Grok 4.6.

这一功能已通过 Gemini API 在 Google AI Studio 上对视频上传和 YouTube 视频开放：https://ai.google.dev/gemini-api/docs/video-understanding#agentic-video-understanding 以及在 Gemini 企业代理平台：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/video-understanding 间提供。开发者在 API 配置中将处理模式设置为 “agentic”，并按标准 Gemini API 令牌费率支付，无需额外费用。更多详情可参见开发者指南：https://aistudio.google.com/learn/agentic-video-understanding-with-gemini。广告

Google 计划将这些改进带入自家产品。该功能很快将在 Flash 和 Flash Lite 设备的所有 Gemini 应用用户中推出。在未来几个月中，基于代理的视频分析还将为播放页面上的“Ask YouTube” 功能提供支持：https://the-decoder.com/googles-ask-youtube-turns-video-search-into-a-conversation/，使答案更紧密地对应视频中实际可见的内容。

保持对人工智能的了解。清晰、有用，无废话。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

解码器：https://the-decoder.com/

Google is adding agent-based video analysis to several Gemini models. Instead of scanning a video frame by frame at a fixed rate, the model hunts for relevant sections on its own, which Google says cuts token usage and costs by a wide margin.

The latest models, Gemini 3.7 Flash：https://the-decoder.com/gemini-3-7-flash-lands-with-coding-gains-and-undercuts-its-three-week-old-predecessors-price-by-50/, 3.6 Flash, and 3.5 Flash-Lite：https://the-decoder.com/googles-gemini-3-5-flash-follows-anthropic-and-openai-in-making-newer-ai-models-significantly-pricier/, can pick up moments shorter than one second, including state changes or cuts that would slip through at one frame per second. Google says this makes automated video editing far more precise.

The system can also track down individual scenes in hours of footage without burning through millions of tokens. It spots anomalies by resampling suspicious time windows at a higher frame rate and accurately counts repeated movements and individual objects over time. Ad

Until now, Gemini relied on static processing, sampling video at a fixed frame rate of one frame per second by default and adjustable through the API. Since native video analysis launched in 2025：https://the-decoder.com/googles-gemini-models-add-native-video-understanding/, Gemini has been transcribing the audio track and analyzing frames on a per-second basis. Ad

Google says the agent-based variant ties the model's reasoning directly to native video tools. The model decides on its own which sections to look at, at what speed, and through which modality, whether that means frames, audio, or transcript. It only pulls the moments and signals it actually needs for a given task.

Gemini now uses an internal tool to grab just the relevant portion of the video file. Developers could build this kind of selective approach manually before, but the model now handles it on its own. Ad

The approach builds on "agentic vision," which Google shipped for Gemini 3 Flash in January：https://the-decoder.com/google-deepmind-gives-gemini-3-flash-the-ability-to-actively-explore-images-through-code/. That feature let the model write and run Python code to zoom, crop, and annotate images, checking each result in a think-act-observe loop before responding. It didn't work automatically in every case at launch, but the groundwork was already there. When Google announced Gemini 3 Flash back in December：https://the-decoder.com/google-makes-gemini-3-flash-the-default-for-search-and-slashes-reasoning-costs/, the company flagged visual and spatial reasoning for video as a coming capability.

The efficiency gains show up most with long videos, anywhere from 10-minute tutorials to 90-minute lectures and multi-hour recordings. With static processing, developers had to pick between high token costs and methods that throw away important details. On Google's own benchmarks, including LongVideoBench, Gemini 3.7 Flash with agent-based analysis scores the highest overall quality and delivers the best mix of accuracy and cost efficiency. Ad

The feature is live for video uploads and YouTube videos through the Gemini API in Google AI Studio：https://ai.google.dev/gemini-api/docs/video-understanding#agentic-video-understanding and on the Gemini Enterprise Agent Platform：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/video-understanding. Developers set the processing mode to "agentic" in the API config and pay standard Gemini API token rates with no added fee. More details are in the Developer Guide：https://aistudio.google.com/learn/agentic-video-understanding-with-gemini. Ad

Google plans to bring these improvements to its own products too. The feature should roll out soon to all Gemini app users on Flash and Flash Lite devices. Over the coming months, agent-based video analysis will also power the "Ask YouTube"：https://the-decoder.com/googles-ask-youtube-turns-video-search-into-a-conversation/ feature on the playback page, giving answers more closely tied to what's actually visible in the video.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/

情报判断

Aioga 编辑摘要

Google 为 Gemini 3.7 Flash、3.6 Flash 和 3.5 Flash-Lite 加入智能体式视频分析。模型可自主选择相关片段、帧率及音频、画面或转录等信息来源，标题称 token 用量最多可降低 88%。

背景分析

此前 Gemini 默认按每秒一帧进行静态视频处理，并结合音频转录分析内容。此次方案由模型自行定位所需时间段，可识别短于一秒的状态变化或剪辑，并对可疑片段提高采样率。

Aioga 观点

Aioga 判断：这项变化的核心不只是减少采样，而是让视频分析从固定规则转向按任务主动取证。若来源所述效果稳定，长视频中的细节保留与成本控制可能获得更好的平衡。

影响与后续

可能影响：长教程、讲座和多小时录制内容的分析方式可能改变；开发者需要重新评估固定采样方案与模型自主取样的准确性、成本及可控性，单项基准结果不足以代表所有场景。后续观察：应关注该能力在不同长度、类型和任务中的实际表现，尤其是短暂事件识别、重复动作计数、场景检索，以及来源所述质量与成本优势能否在更多测试中复现。