广告
加载中

谷歌Gemini上线智能体视频分析 最高削减88%token用量

亿邦AI 2026-09-03 10:25
亿邦AI 2026/09/03 10:25

邦小白快读

EN
全文速览

本文核心介绍谷歌推出的Gemini Flash系列大模型全新智能体驱动视频分析功能,普通用户可关注以下干货内容

1. 功能核心升级:原来的原生视频分析是固定每秒1帧采样,所有内容全量处理,新功能改为由模型自主判断需要分析的片段、采样速度,只提取任务需要的信息,无需手动搭建分析流程,使用更便捷。

2. 核心优势明确:官方数据显示新功能最高可削减88%的token用量,降低66%使用成本,同时提升分析准确率,还能捕捉短于1秒的画面片段,不会遗漏转场等关键变化,支持长视频场景定位单个场景,识别异常。

3. 使用方式清晰:目前该功能已经通过API开放,无需额外付费,后续很快会推送给所有Gemini app的Flash系列用户,还将接入YouTube的询问功能,普通用户很快就能体验。

该新功能可为品牌的视频内容营销、内容复盘分析降本提效,带来多方面新机会,核心干货如下

1. 成本优势明显:品牌日常需要处理大量直播回放、种草素材、品牌推广视频,原有静态分析模式只能在高成本和丢失关键细节中二选一,新功能最高降低66%的使用成本,解决了这一核心痛点,适合品牌做海量内容处理。

2. 功能适配品牌需求:新功能可在数小时的视频素材中快速定位品牌露出、产品展示、用户互动等特定场景,还能准确统计相关动作次数,帮助品牌快速复盘营销内容的效果,大幅提升自动视频编辑的精度,提高内容产出效率。

3. 生态对接带来营销便利:该功能后续将接入YouTube的询问功能,品牌可以依托该功能优化YouTube频道的内容互动体验,提升品牌内容的触达效果,强化用户粘性。

该新功能为从事视频内容运营、出海业务的卖家提供了降本增效的新工具,带来明确的业务机会,核心干货如下

1. 适配卖家降本需求:卖家日常需要处理大量直播回放、产品教程、种草短视频等内容,原有分析方案成本高,中小卖家难以负担,新功能最高降低66%的使用成本,同时不会丢失关键细节,完美适配卖家控制运营成本的需求。

2. 功能满足卖家业务场景:新功能可以快速在数小时直播素材中定位产品讲解、用户提问等特定场景,准确统计产品出镜次数、用户互动动作,帮助卖家快速完成直播复盘,提升短视频剪辑制作效率,优化内容运营效果。

3. 接入门槛低:目前该功能已经通过Gemini API开放,按原有token费率付费,没有额外费用,后续会推送给普通用户,卖家可以低门槛接入试用,不用额外付出成本就能优化自身的视频内容运营流程。

该新技术为工厂推进数字化、智能化升级提供了新的启示和商业机会,核心干货内容如下

1. 技术思路带来数字化升级启示:该功能采用智能体自主判断、按需采样提取信息的思路,打破了原有固定采样的模式,这一思路可以直接借鉴到工厂的工业监控视频分析、生产流程智能化改造中,解决原有工厂监控分析固定采样成本高、易漏异常的问题。

2. 可直接应用于工厂生产场景:工厂做生产流程优化时,需要统计重复生产动作数量、工件数量,识别生产异常,依托该功能可以低成本完成这些需求,还能在长时间监控素材中快速定位异常窗口,提升异常识别准确率,帮助工厂实现低成本的生产智能监控。

3. 指明低成本智能化方向:该功能通过大模型能力按需调用资源的方式,大幅降低了智能化应用的成本,给工厂做自有智能化应用开发提供了参考,帮助工厂降低智能化改造的成本压力,更快推进数字化转型。

谷歌推出的该新功能,给AI视频分析服务商指明了行业发展方向,也提供了成熟的技术参考,核心干货如下

1. 明确了客户核心痛点:当前长视频分析需求快速增长,原有静态固定采样方案始终无法解决高token成本和丢失关键细节不可兼得的痛点,这是服务商开发新产品可以切入的核心方向。

2. 提供了成熟的技术方案参考:该功能采用智能体自主决策,按需采样提取信息的技术路线,官方测试数据显示,该方案最高可削减88%token用量,降低66%使用成本,同时提升分析准确率,在通用长视频理解测试中拿到最高得分,技术可行性已经得到验证,可给服务商开发同类产品提供参考。

3. 明确了可落地的应用场景:该功能已经验证了在长视频内容分析、自动视频编辑、异常场景识别、特定场景定位等多个场景的可用性,服务商可以基于该技术思路,开发面向不同行业的定制化视频分析解决方案,拓展自身的业务空间。

谷歌推出该新功能的思路和做法,给大模型平台的功能迭代、运营推广提供了多方面参考,核心干货如下

1. 明确了平台用户的核心需求:当前大模型平台用户对多模态长视频分析的需求越来越高,核心痛点是需要兼顾分析准确率和使用成本,原有方案无法满足,平台需要针对性优化产品功能,解决用户痛点才能提升用户留存。

2. 提供了新功能推广的运营参考:谷歌将该新功能按原有token费率开放,不收取额外费用,降低了用户的试用门槛,能够快速吸引用户试用新功能,这种运营方式值得各类大模型平台借鉴。

3. 提供了功能迭代和生态协同的参考:该功能依托此前上线的智能视觉能力逐步迭代开发,渐进式推出,降低了技术风险,后续还对接谷歌生态内的YouTube产品,强化生态协同,这种做法可以帮助平台规避新技术推广的风险,同时提升整个生态的价值,值得平台参考。

谷歌推出的智能体驱动视频分析功能,代表了大模型多模态应用领域的新动向,具有较高的研究价值,核心干货如下

1. 指明了产业新的技术动向:此前大模型视频分析领域普遍采用固定帧率静态采样的技术方案,该方案存在难以解决的固有痛点,谷歌推出的智能体自主决策采样的新方案,验证了智能体架构在多模态信息处理领域的优势,是行业新的技术发展方向,值得深入研究。

2. 提供了公开可参考的研究数据:谷歌公开了该方案的明确效果数据,最高可削减88%token用量,降低66%使用成本,同时在行业通用的LongVideoBench长视频理解基准测试中拿到了整体最高得分,这些数据可以为相关领域的研究提供对比样本,支撑后续研究开展。

3. 提供了新技术商业化落地的研究样本:该新功能采用先开放API给开发者测试,后续逐步推送给C端用户、接入生态内产品的路径,采用原有收费标准不额外加价的商业化模式,为同类大模型新技术的商业化落地提供了可研究的样本,具有较高的参考价值。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article focuses on the new agent-driven video analysis function launched by Google for its Gemini Flash large model series, and key takeaways for general readers are as follows:

1. Core functional upgrade: Unlike the original native video analysis which relied on fixed 1-frame-per-second sampling and full-content processing, the new function lets the model autonomously determine which video clips to analyze and adjust sampling rates automatically. It only extracts information required for the task, eliminating the need for manual workflow building and delivering a much more user-friendly experience.

2. Clear core advantages: Official data shows the new function reduces token usage by up to 88% and cuts usage costs by 66%, while improving analysis accuracy. It can also capture video segments shorter than 1 second, avoiding missed key changes such as scene transitions. It supports positioning individual scenes in long videos and identifying anomalies.

3. Straightforward access: The function is already available via API at no additional cost. It will soon roll out to all Gemini Flash users on the Gemini app, and will also be integrated into YouTube's question-answering feature, meaning general users will be able to access it very soon.

This new function can cut costs and improve efficiency for brands' video content marketing and post-campaign performance analysis, unlocking multiple new opportunities. Key takeaways for brands are as follows:

1. Significant cost advantages: Brands process large volumes of content daily, including live stream replays, influencer marketing materials, and branded promotion videos. The original static analysis model forced brands to choose between high costs and lost key details. The new function reduces usage costs by up to 66%, solving this core pain point and making it ideal for processing massive content volumes for brands.

2. Functionality tailored to brand needs: The new function can quickly locate specific scenes such as brand mentions, product displays, and user interactions from hours of video footage, and accurately count related actions. This helps brands quickly review the performance of marketing content, greatly improving the accuracy of automatic video editing and boosting content production efficiency.

3. Ecosystem integration delivers marketing convenience: The function will later be integrated into YouTube's question-answering feature, allowing brands to improve the interactive experience of their YouTube channel content, boost the reach of branded content, and strengthen user engagement.

This new function is a new cost-cutting efficiency-boosting tool for sellers engaged in video content operations and cross-border business, bringing clear business opportunities. Key takeaways for sellers are as follows:

1. Aligns with sellers' cost-reduction needs: Sellers process large volumes of content daily, including live stream replays, product tutorials, and short marketing videos. Original analysis solutions were too costly for small and medium-sized sellers to afford. The new function cuts usage costs by up to 66% without sacrificing key details, perfectly matching sellers' need to control operational costs.

2. Functionality meets sellers' business scenarios: The new function can quickly locate specific scenes such as product explanations and user questions from hours of live stream footage, and accurately count product appearance times and user interactions. This helps sellers complete post-live reviews quickly, improve short video editing efficiency, and optimize content operation performance.

3. Low access barrier: The function is already available via the Gemini API, billed at the original token rate with no extra charges. It will roll out to general users soon, allowing sellers to access and test it with a low barrier, and optimize their video content operation workflows without additional costs.

This new technology brings new insights and business opportunities for factories advancing digital and intelligent transformation. Key takeaways for factories are as follows:

1. Technical insights for digital upgrading: This function adopts an agent-driven approach of autonomous judgment and on-demand sampling to extract information, breaking the traditional fixed sampling model. This approach can be directly adapted to factories' industrial surveillance video analysis and intelligent production process transformation, solving the original problems of high costs from fixed sampling and frequent missed anomalies.

2. Directly applicable to factory production scenarios: When optimizing production processes, factories need to count repeated production actions and workpieces, and identify production anomalies. This function can meet these needs at a low cost, and quickly locate anomaly windows from long hours of surveillance footage, improving anomaly detection accuracy and helping factories achieve low-cost intelligent production monitoring.

3. Points to a low-cost path to intelligent transformation: By leveraging large model capabilities to call resources on demand, this function greatly reduces the cost of intelligent applications. It provides a reference for factories developing their own intelligent applications, helping factories reduce cost pressure for transformation and speed up digital transformation progress.

Google's new function points to the future development direction for the AI video analysis service industry and provides a mature technical reference. Key takeaways for service providers are as follows:

1. Clarifies customers' core pain points: Demand for long video analysis is growing rapidly, and traditional static fixed sampling solutions have never been able to solve the core tradeoff between high token costs and lost key details, which is the core gap service providers can target when developing new products.

2. Provides a mature reference technical solution: This function adopts the technical route of agent-driven autonomous decision-making and on-demand information sampling. Official test data shows the solution reduces token usage by up to 88% and cuts costs by 66%, while improving analysis accuracy, and it achieved the top overall score on the general long video understanding benchmark. The technical feasibility is already proven, providing a reference for service providers developing similar products.

3. Identifies actionable application scenarios: This function has already verified usability across multiple scenarios including long-form video analysis, automatic video editing, anomaly detection, and specific scene positioning. Service providers can develop customized video analysis solutions for different industries based on this technical approach to expand their business scope.

Google's approach in launching this new function provides multiple references for large model platforms on product iteration and operation promotion. Key takeaways for large model platforms are as follows:

1. Clarifies core user needs: Users of large model platforms have growing demand for multimodal long video analysis, with the core pain point of requiring both high analysis accuracy and low usage costs, which traditional solutions cannot meet. Platforms need to optimize product functions targeting this pain point to improve user retention.

2. Provides an operational reference for new function promotion: Google opened the new function at the original token rate with no extra charges, lowering users' trial barrier and quickly attracting users to test the new function. This operational approach is worth learning from for all large model platforms.

3. Provides a reference for function iteration and ecosystem collaboration: This function was developed iteratively based on Google's previously launched intelligent vision capabilities and launched incrementally, reducing technical risk. It will later be integrated into YouTube, Google's in-ecosystem product, to strengthen ecosystem synergy. This approach helps platforms avoid risks when rolling out new technologies while increasing value for the entire ecosystem, and is worthy of reference for platforms.

Google's agent-driven video analysis function represents a new trend in large model multimodal applications and carries high research value. Key takeaways for researchers are as follows:

1. Identifies a new industrial technical direction: Previously, the large model video analysis field generally adopted a static fixed-frame-rate sampling technical solution, which has inherent unsolved pain points. Google's new solution of agent-driven autonomous sampling decision-making verifies the advantages of agent architecture in multimodal information processing, marking a new technical development direction for the industry worthy of in-depth research.

2. Provides publicly available reference research data: Google has released clear performance data for this solution: it reduces token usage by up to 88% and cuts costs by 66%, and it achieved the top overall score on the industry-standard LongVideoBench benchmark for long video understanding. These data provide comparison samples for research in related fields to support subsequent research work.

3. Provides a research sample for new technology commercialization: This new function follows the path of first opening API access for developer testing, then gradually rolling out to end consumers and integrating into in-ecosystem products, and adopts a commercial model of keeping original pricing without extra charges. It provides a researchable sample for the commercialization of similar large model new technologies with high reference value.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年9月2日,谷歌为旗下Gemini Flash系列多模态大模型配备智能体驱动的视频分析功能。Gemini是Google Deepmind推出的通用多模态大模型,具备文本、图像、视频、音频、代码跨模态推理能力,Flash系列为该模型线中兼顾性能与成本的产品线。

此前Gemini的原生视频分析采用静态处理模式,默认以每秒1帧的固定帧率采样视频,逐秒转录音频、分析画面,相关功能2025年正式上线。新的智能体视频分析取消固定采样逻辑,由模型自主判断需要分析的视频片段、采样速度及调用模态,可选择画面、音频或字幕任意组合,仅提取任务所需的片段及信号,无需开发者手动搭建选择性分析流程。

官方公开信息显示,新方案最高可削减88%的token用量,降低66%的使用成本,同时提升分析准确率。目前Gemini 3.7 Flash、3.6 Flash、3.5 Flash-Lite三款模型已支持该功能,可捕捉短于1秒的画面片段,覆盖固定帧率采样会遗漏的状态变化、画面转场,大幅提升自动视频编辑精度。系统还可在数小时的视频素材中定位单个场景,对可疑时间窗口提高采样帧率识别异常,准确统计重复动作及单个物体数量。

该功能基于2026年1月上线的Gemini 3 Flash智能视觉能力开发,此前该功能支持模型编写运行Python代码,对图片进行缩放、裁剪、标注操作,在思考行动观察循环中校验结果。2025年12月Gemini 3 Flash发布时,谷歌就预告视频视觉与空间推理能力将在后续推出。

效率提升在长视频场景最为明显,覆盖10分钟教程、90分钟讲座到数小时录制素材等多种场景。静态处理模式下,开发者只能在高token成本与丢失重要细节之间二选一。LongVideoBench是行业通用的长视频理解基准测试,包含3763条最长1小时的多主题视频及6678道人工标注的选择题,谷歌自有测试数据显示,搭载智能体分析功能的Gemini 3.7 Flash在该基准中整体质量得分最高,兼顾准确率与成本效率。

目前该功能已通过Gemini API开放,支持视频上传及YouTube视频的特定场景查找,开发者在Google AI Studio和Gemini企业智能体平台使用时,在API配置中将处理模式设为智能体即可,按标准Gemini API token费率付费,无额外费用。后续该功能将很快推送给所有Gemini app的Flash及Flash Lite设备用户,未来几个月还将接入YouTube播放页的询问YouTube功能,输出答案将更贴合视频实际画面内容。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

谷歌Gemini智能体视频分析功能有什么优势?

该功能取消固定采样逻辑,由模型自主判断需分析的视频片段、采样速度及调用模态,最高可削减88%的token用量,降低66%的使用成本,同时提升分析准确率,可捕捉短于1秒的画面片段,覆盖固定帧率采样会遗漏的状态变化、画面转场。

哪些Gemini模型支持智能体视频分析功能?

目前Gemini 3.7 Flash、3.6 Flash、3.5 Flash-Lite三款模型已支持该功能,该功能已通过Gemini API开放,开发者可在Google AI Studio和Gemini企业智能体平台配置使用,后续将推送给所有Gemini app的Flash及Flash Lite设备用户。

Gemini智能体视频分析功能适用于哪些场景?

该功能效率提升在长视频场景最为明显,覆盖10分钟教程、90分钟讲座到数小时录制素材等多种场景,可在数小时的视频素材中定位单个场景,对可疑时间窗口提高采样帧率识别异常,准确统计重复动作及单个物体数量。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0