Google’s latest AI model uses a web browser like you do

来源：The Verge | 发布时间：2025-10-31

Google is previewing a new Gemini AI model designed to navigate and interact with the web via a browser, letting AI agents do things inside interfaces designed for use by people and not robots. The model, called Gemini 2.5 Computer Use, uses “visual understanding and reasoning capabilities” to analyze a user’s request and carry out a task, such as filling out and submitting a form.

It can be used for UI testing or navigating interfaces made for people who don’t have an API or other direct connection available. Other versions of this model have been used for agentic features in AI Mode and Project Mariner, a research prototype that uses AI agents to carry out tasks on its own in a browser, like adding items to your cart based on a list of ingredients.

Google’s announcement comes just one day after OpenAI revealed new apps for ChatGPT as part of its annual Dev Day, and continues to focus its attention on its ChatGPT Agent feature that can complete complex tasks on your behalf. Meanwhile, Anthropic had already released a version of its Claude AI model with “computer use” last year.

Google posted some demo videos showing its computer use tool in action, and notes that they are sped up 3x.

Google says its computer use model “outperforms leading alternatives on multiple web and mobile benchmarks.” Unlike ChatGPT Agent and Anthropic’s computer use tool, Google’s new AI model only has access to a browser — not an entire computer environment. Google notes that it shows “it is not yet optimized for desktop OS-level control” and currently supports 13 actions, including opening a web browser, typing text, as well as dragging and dropping elements.

Gemini 2.5 Computer Use is available to developers through Google AI Studio and Vertex AI, but there’s also a demo on Browserbase, where you watch as it completes tasks, like “Play a game of 2048” or “Browse Hacker News for trending debates.”

谷歌发布Gemini 2.5 Computer Use模型：专攻浏览器交互

IT之家 10 月 8 日消息，谷歌正在预览一款全新的 Gemini 人工智能模型，该模型旨在通过浏览器浏览网络并与之交互，使 AI 智能体能够在原本为人而非机器人设计的用户界面中执行操作。这款名为“Gemini 2.5 计算机使用”（Gemini 2.5 Computer Use）的模型，利用“视觉理解与推理能力”分析用户的请求，并完成相应任务，例如填写并提交表单。

该模型可用于用户界面测试，或用于操作那些仅面向人类用户、未提供 API 或其他直接接口的系统。此前，此类模型已应用于谷歌 AI 模式（AI Mode）中的智能体功能，以及研究原型项目“Mariner”，该项目利用 AI 智能体在浏览器中自主执行任务，例如根据食材清单自动将商品添加到购物车。

谷歌此次发布恰逢 OpenAI 在其年度开发者日（Dev Day）宣布推出 ChatGPT 新应用的次日。OpenAI 正持续聚焦其“ChatGPT 智能体”（ChatGPT Agent）功能，该功能可代表用户完成复杂任务。与此同时，Anthropic 公司已于去年发布了其 Claude AI 模型的“计算机使用”（computer use）版本。

IT之家注意到，谷歌发布了一些演示视频，展示了“计算机使用”工具的实际运行效果，并注明视频播放速度已加快至 3 倍。谷歌表示，其计算机使用模型“在多个网页和移动端基准测试中优于现有领先方案”。与 ChatGPT 智能体和 Anthropic 的计算机使用工具不同，谷歌的新 AI 模型仅限于访问浏览器环境，无法操控整个计算机系统。谷歌特别指出，该模型“尚未针对桌面操作系统层级的控制进行优化”，目前支持 13 种操作，包括打开网页浏览器、输入文本，以及拖放页面元素等。

Gemini 2.5 计算机使用模型现已通过 Google AI Studio 和 Vertex AI 向开发者开放。此外，Browserbase 平台也提供了在线演示，用户可实时观看该模型完成诸如“玩一局 2048 游戏”或“浏览 Hacker News 上的热门讨论”等任务。