January 23, 2025 1月 23， 2025
Introducing Operator Operator 简介
A research preview of an agent that can use its own browser to perform tasks for you. Available to Pro users in the U.S.
代理的研究预览，该代理可以使用自己的浏览器为您执行任务。适用于美国的 Pro 用户
Go to Operator 转到运算符(opens in a new window)
Today we’re releasing Operator(opens in a new window), an agent that can go to the web to perform tasks for you. Using its own browser, it can look at a webpage and interact with it by typing, clicking, and scrolling. It is currently a research preview , meaning it has limitations and will evolve based on user feedback. Operator is one of our first agents, which are AIs capable of doing work for you independently—you give it a task and it will execute it.
今天，我们发布了 Operator，一个可以转到 Web 为您执行任务的代理。使用自己的浏览器，它可以查看网页并通过键入、单击和滚动与之交互。它目前是一个研究预览版 ，这意味着它有局限性，并将根据用户反馈进行改进。Operator 是我们的第一批代理之一，它们是能够独立为您完成工作的 AI，您给它一个任务，它就会执行它。
Operator can be asked to handle a wide variety of repetitive browser tasks such as filling out forms, ordering groceries, and even creating memes. The ability to use the same interfaces and tools that humans interact with on a daily basis broadens the utility of AI, helping people save time on everyday tasks while opening up new engagement opportunities for businesses.
可以要求 Operator 处理各种重复的浏览器任务，例如填写表格、订购杂货，甚至创建模因。使用人类日常交互的相同界面和工具的能力扩大了 AI 的实用性，帮助人们节省日常任务的时间，同时为企业开辟新的参与机会。
To ensure a safe and iterative rollout, we are starting small. Starting today, Operator is available to Pro users in the U.S. at operator.chatgpt.com(opens in a new window). This research preview allows us to learn from our users and the broader ecosystem, refining and improving as we go. Our plan is to expand to Plus, Team, and Enterprise users and integrate these capabilities into ChatGPT in the future.
为了确保安全和迭代的推出，我们从小处着手。从今天开始，Operator 以 operator.chatgpt.com 的价格向美国的 Pro 用户提供。此研究预览使我们能够从用户和更广泛的生态系统中学习，并在过程中进行完善和改进。我们的计划是扩展到 Plus、Team 和 Enterprise 用户，并将这些功能在未来集成到 ChatGPT 中。
How Operator works Operator 的工作原理
Operator is powered by a new model called Computer-Using Agent (CUA). Combining GPT-4o’s vision capabilities with advanced reasoning through reinforcement learning, CUA is trained to interact with graphical user interfaces (GUIs)—the buttons, menus, and text fields people see on a screen.
Operator 由一种称为 Computer-Using Agent （CUA） 的新模型提供支持。CUA 通过强化学习将 GPT-4o 的视觉功能与高级推理相结合，经过训练可以与图形用户界面 （GUI） 交互，即人们在屏幕上看到的按钮、菜单和文本字段。
Operator can “see” (through screenshots) and “interact” (using all the actions a mouse and keyboard allow) with a browser, enabling it to take action on the web without requiring custom API integrations.
操作员可以与浏览器“查看”（通过屏幕截图）和“交互”（使用鼠标和键盘允许的所有操作），使其能够在 Web 上执行操作，而无需自定义 API 集成。
If it encounters challenges or makes mistakes, Operator can leverage its reasoning capabilities to self-correct. When it gets stuck and needs assistance, it simply hands control back to the user, ensuring a smooth and collaborative experience.
如果遇到挑战或犯错，Operator 可以利用其推理能力进行自我纠正。当它卡住并需要帮助时，它只需将控制权交还给用户，确保流畅的协作体验。
While CUA is still in early stages and has limitations, it sets new state-of-the-art benchmark results in WebArena and WebVoyager, two key browser use benchmarks. Read more about evals and the research behind Operator in our research blog post.
虽然 CUA 仍处于早期阶段并且存在局限性，但它在 WebArena 和 WebVoyager 这两个关键的浏览器使用基准测试中设定了新的最先进的基准测试结果。在我们的研究博客文章中阅读有关 evals 和 Operator 背后的研究的更多信息。
How to use 如何使用
To get started, simply describe the task you’d like done and Operator can handle the rest. Users can choose to take over control of the remote browser at any point, and Operator is trained to proactively ask the user to take over for tasks that require login, payment details, or when solving CAPTCHAs.
要开始使用，只需描述您要完成的任务，Operator 可以处理其余的工作。用户可以随时选择接管远程浏览器的控制权，并且 Operator 经过培训，可以主动要求用户接管需要登录、付款详细信息或解决 CAPTCHA 的任务。
Users can personalize their workflows in Operator by adding custom instructions, either for all sites or for specific ones, such as setting preferences for airlines on Booking.com. Operator lets users save prompts for quick access on the homepage, ideal for repeated tasks like restocking groceries on Instacart. Similar to using multiple tabs on a browser, users can have Operator run multiple tasks simultaneously by creating new conversations, like ordering a personalized enamel mug on Etsy while booking a campsite on Hipcamp.
用户可以通过为所有站点或特定站点添加自定义说明，例如为 Booking.com 上的航空公司设置首选项，在 Operator 中个性化其工作流。Operator 允许用户保存提示以便在主页上快速访问，非常适合重复任务，例如在 Instacart 上补货。与在浏览器上使用多个选项卡类似，用户可以通过创建新对话让 Operator 同时运行多个任务，例如在 Hipcamp 上预订露营地时在 Etsy 上订购个性化的珐琅杯。
Customizing 定制Optimizing 优化
Customizing 定制Optimizing 优化
Ecosystem & users 生态系统和用户
Operator(opens in a new window) transforms AI from a passive tool to an active participant in the digital ecosystem. It will streamline tasks for users and bring the benefits of agents to companies that want innovative customer experiences and desire higher rates of conversion. We’re collaborating with companies like DoorDash, Instacart, OpenTable, Priceline, StubHub, Thumbtack, Uber, and others to ensure Operator addresses real-world needs while respecting established norms. In addition to these collaborations, we see a lot of potential to improve the accessibility and efficiency of certain workflows, particularly in public sector applications. To explore these use cases further, we’re working with organizations like the City of Stockton(opens in a new window) to make it easier to enroll in city services and programs.
Operator 将 AI 从被动工具转变为数字生态系统的主动参与者。它将简化用户的任务，并为需要创新客户体验和更高转化率的公司带来代理的好处。我们正在与 DoorDash、Instacart、OpenTable、Priceline、StubHub、Thumbtack、Uber 等公司合作，以确保 Operator 在遵守既定规范的同时满足实际需求。除了这些合作之外，我们还看到了提高某些工作流程的可访问性和效率的巨大潜力，尤其是在公共部门应用程序中。为了进一步探索这些使用案例，我们正在与斯托克顿市等组织合作，以便更轻松地注册城市服务和计划。
“As we learn more about Operator during its research preview, we’ll be better equipped to identify ways that AI can make civic engagement even easier for our residents.”
“随着我们在 Operator 的研究预览期间对 Operator 的了解越来越多，我们将能够更好地确定 AI 如何使我们的居民更容易进行公民参与。”
Jamil Niazi, Director of Information Technology at City of Stockton
Jamil Niazi，斯托克顿市信息技术总监
By releasing Operator to a limited audience initially, we aim to learn quickly and refine its capabilities based on real-world feedback, ensuring we balance innovation with trust and safety. This collaborative approach helps ensure Operator delivers meaningful value to users, creators, businesses, and public sector organizations alike.
通过最初向有限的受众发布 Operator，我们的目标是快速学习并根据真实世界的反馈完善其功能，确保我们在创新与信任和安全之间取得平衡。这种协作方法有助于确保 Operator 为用户、创作者、企业和公共部门组织提供有意义的价值。
“OpenAI’s Operator is a technological breakthrough that makes processes like ordering groceries incredibly easy.”
“OpenAI 的 Operator 是一项技术突破，它使订购杂货等流程变得异常简单。”
Daniel Danker, Chief Product Officer at Instacart
Daniel Danker，Instacart 首席产品官
Safety and privacy 安全和隐私
Ensuring Operator is safe to use is a top priority, with three layers of safeguards to prevent abuse and ensure users are firmly in control.
确保 Operator 可以安全使用是重中之重，有三层保护措施来防止滥用并确保用户牢牢控制。
First, Operator is trained to ensure that the person using it is always in control and asks for input at critical points.
首先，操作员经过培训，以确保使用它的人始终处于控制之中，并在关键点要求输入。
-
Takeover mode: Operator asks the user to take over when inputting sensitive information into the browser, such as login credentials or payment information. When in takeover mode, Operator does not collect or screenshot information entered by the user.
接管模式： 操作员在浏览器中输入敏感信息（如登录凭证或支付信息）时要求用户接管。在接管模式下，Operator 不会收集用户输入的信息或对其进行屏幕截图。 -
User confirmations : Before finalizing any significant action, such as submitting an order or sending an email, Operator should ask for approval.
用户确认 ：在完成任何重大操作（例如提交订单或发送电子邮件）之前，操作员应请求批准。 -
Task limitations: Operator is trained to decline certain sensitive tasks, such as banking transactions or those requiring high-stakes decisions, like making a decision on a job application.
任务限制： 操作员接受过拒绝某些敏感任务的培训，例如银行交易或需要高风险决策的任务，例如对工作申请做出决定。 -
Watch mode : On particularly sensitive sites, such as email or financial services, Operator requires close supervision of its actions, allowing users to directly catch any potential mistakes.
监视模式 ：在特别敏感的站点（例如电子邮件或金融服务）上，Operator 需要密切监督其操作，使用户能够直接捕获任何潜在的错误。
Next, we’ve made it easy to manage data privacy in Operator.
接下来，我们简化了在 Operator 中管理数据隐私的过程。
-
Training opt out: Turning off ‘Improve the model for everyone’ in ChatGPT settings means data in Operator will also not be used to train our models.
训练选择退出： 在 ChatGPT 设置中关闭“为每个人改进模型”意味着 Operator 中的数据也不会用于训练我们的模型。 -
Transparent data management: Users can delete all browsing data and log out of all sites with one click under the Privacy section of Operator settings. Past conversations in Operator can also be deleted with one click.
透明的数据管理： 用户可以在 Operator 设置的 Privacy 部分下一键删除所有浏览数据并注销所有站点。Operator 中过去的对话也可以一键删除。
Lastly, we’ve built defenses against adversarial websites that may try to mislead Operator through hidden prompts, malicious code, or phishing attempts:
最后，我们建立了防御措施，抵御可能试图通过隐藏提示、恶意代码或网络钓鱼尝试来误导 Operator 的敌对网站：
-
Cautious navigation : Operator is designed to detect and ignore prompt injections.
谨慎导航 ：Operator 旨在检测和忽略提示注入。 -
Monitoring : A dedicated “monitor model” watches for suspicious behavior and can pause the task if something seems off.
监控 ：专用的“监控模型”会监视可疑行为，如果出现问题，可以暂停任务。 -
Detection pipeline : Automated and human review processes continuously identify new threats and quickly update safeguards.
检测管道 ：自动化和人工审查流程持续识别新威胁并快速更新保护措施。
We know bad actors may try to misuse this technology. That’s why we’ve designed Operator to refuse harmful requests and block disallowed content. Our moderation systems can issue warnings or even revoke access for repeated violations, and we’ve integrated additional review processes to detect and address misuse. We’re also providing guidance on how to interact with Operator in compliance with our Usage Policies.
我们知道，不良行为者可能会试图滥用这项技术。因此，我们将 Operator 设计为拒绝有害请求并阻止不允许的内容。我们的审核系统可以对重复的违规行为发出警告甚至撤销访问权限，并且我们集成了额外的审核流程来检测和解决滥用问题。我们还将提供有关 如何根据我们的使用政策与 Operator 互动的指导。
While Operator is designed with these safeguards, no system is flawless and this is still a research preview; we are committed to continuous improvement through real-world feedback and rigorous testing. For more on our approach, visit the safety section of the Operator research blog.
虽然 Operator 在设计时就采用了这些保护措施，但没有一个系统是完美的，这仍然是一个研究预览;我们致力于通过真实世界的反馈和严格的测试不断改进。有关我们方法的更多信息，请访问 Operator 研究博客的安全部分。
Limitations 局限性
Operator is currently in an early research preview, and while it’s already capable of handling a wide range of tasks, it’s still learning, evolving and may make mistakes. For instance, it currently encounters challenges with complex interfaces like creating slideshows or managing calendars. Early user feedback will play a vital role in enhancing its accuracy, reliability, and safety, helping us make Operator better for everyone.
Operator 目前处于早期研究预览阶段，虽然它已经能够处理各种任务，但它仍在学习和发展，并且可能会犯错误。例如，它目前在创建幻灯片或管理日历等复杂界面方面遇到了挑战。早期用户反馈将在提高其准确性、可靠性和安全性方面发挥至关重要的作用，帮助我们让 Operator 更好地为每个人服务。
What’s next 下一步
CUA in the API: We plan to expose the model powering Operator, CUA, in the API soon so that developers can use it to build their own computer-using agents.
**API 中的 CUA：**我们计划很快在 API 中公开为 Operator 提供支持的模型 CUA，以便开发人员可以使用它来构建自己的计算机使用代理。
Enhanced Capabilities: We’ll continue to improve Operator’s ability to handle longer and more complex workflows.
增强功能： 我们将继续提高 Operator 处理更长、更复杂的工作流程的能力。
Wider Access: We plan to expand Operator(opens in a new window) to Plus, Team, and Enterprise users and integrate its capabilities directly into ChatGPT in the future once we are confident in its safety and usability at scale, unlocking seamless real-time and asynchronous task execution.
更广泛的访问： 一旦我们对 ChatGPT 的大规模安全性和可用性有信心，我们计划将 Operator 扩展到 Plus、Team 和 Enterprise 用户，并在未来将其功能直接集成到 ChatGPT 中，从而解锁无缝的实时和异步任务执行。
Watch the replay 观看重播
Authors 作者
Foundational research contributors
基础研究贡献者
Casey Chu, David Medina, Hyeonwoo Noh, Noah Jorgensen, Reiichiro Nakano, Sarah Yoo
Casey Chu， David Medina， Hyeonwoo Noh， Noah Jorgensen， Reiichiro Nakano， Sarah Yoo
Core 核心
Andrew Howell, Aaron Schlesinger, Baishen Xu, Ben Newhouse, Bobby Stocker, Devashish Tyagi, Dibyo Majumdar, Eugenio Panero, Fereshte Khani, Geoffrey Iyer, Jiahui Yu, Nick Fiacco, Patrick Goethe, Sam Jau, Shunyu Yao, Stephan Casas, Yash Kumar, Yilong Qin
Andrew Howell， Aaron Schlesinger， Baishen Xu， Ben Newhouse， Bobby Stocker， Devashish Tyagi， Dibyo Majumdar， Eugenio Panero， Fereshte Khani， Geoffrey Iyer， Jiahui Yu， Nick Fiacco， Patrick Goethe， Sam Jau， Shunyu Yao， Stephan Casas， Yash Kumar， 秦一龙
XFN Contributors XFN 贡献者
Abby Fanlo Susk, Aleah Houze, Alex Beutel, Alexander Prokofiev, Andrea Vallone, Andrea Chan, Christina Lim, Derek Chen, Duke Kim, Grace Zhao, Heather Whitney, Houda Nait El Barj, Jake Brill, Jeremy Fine, Joe Fireman, Kelly Stirman, Lauren Yang, Lindsay McCallum, Leo Liu, Mike Starr, Minnia Feng, Mostafa Rohaninejad, Oleg Boiko, Owen Campbell-Moore, Paul Ashbourne, Stephen Imm, Taylor Gordon, Tina Sriskandarajah, Winston Howes
Abby Fanlo Susk， Aleah Houze， Alex Beutel， Alexander Prokofiev， Andrea Vallone， Andrea Chan， Christina Lim， Derek Chen， Duke Kim， Grace Zhao， Heather Whitney， Houda Nait El Barj， Jake Brill， Jeremy Fine， Joe Fireman， Kelly Stirman， Lauren Yang， Lindsay McCallum， Leo Liu， Mike Starr， Minnia Feng， Mostafa Rohaninejad， Oleg Boiko， Owen Campbell-Moore， Paul Ashbourne， Stephen Imm， Taylor Gordon， 蒂娜·斯里坎达拉贾，温斯顿·豪斯
Leads 导致
Aaron Schlesinger (Infrastructure), Casey Chu (Safety and Model Readiness), David Medina (Research Infrastructure), Hyeonwoo Noh (Overall Research), Reiichiro Nakano (Overall Research), Yash Kumar
Aaron Schlesinger（基础设施）、Casey Chu（安全和模型准备）、David Medina（研究基础设施）、Hyeonwoo Noh（整体研究）、Reiichiro Nakano（整体研究）、Yash Kumar
Contributors 贡献
Adam Brandon, Adam Koppel, Adele Li, Ahmed El-Kishky, Akila Welihinda, Alex Karpenko, Alex Nawar, Alex Tachard Passos, Amelia Liu, Andrei Gheorghe, Andrew Duberstein, Andrey Mishchenko, Angela Baek, Ankush Agarwal, Anting Shen, Antoni Baum, Ari Seff, Ashley Tyra, Behrooz Ghorbani, Bo Xu, Brandon McKinzie, Bryan Brandow, Carolina Paz, Cary Hudson, Chak Li, Chelsea Voss, Chen Shen, Chris Koch, Christian Gibson, Christina Kim, Christine McLeavey, Claudia Fischer, Cory Decareaux, Daniel Jacobowitz, Daniel Wolf, David Kjelkerud, David Li, Ehsan Asdar, Elaine Kim, Emilee Goo, Eric Antonow, Eric Hunter, Eric Wallace, Felipe Torres, Fotis Chantzis, Freddie Sulit, Giambattista Parascandolo, Hadi Salman, Haiming Bao, Haoyu Wang, Henry Aspegren, Hyung Won Chung, Ian O’Connell, Ian Sohl, Isabella Fulford, Jake McNeil, James Donovan, Jamie Kiros, Jason Ai, Jason Fedor, Jason Wei, Jay Dixit, Jeffrey Han, Jeffrey Sabin-Matsumoto, Jennifer Griffith-Delgado, Jeramy Han, Jeremiah Currier, Ji Lin, Jiajia Han, Jiaming Zhang, Jiayi Weng, Jieqi Yu, Joanne Jang, Joyce Ruffell, Kai Chen, Kai Xiao, Kevin Button, Kevin King, Kevin Liu, Kristian Georgiev, Kyle Miller, Lama Ahmad, Laurance Fauconnet, Leonard Bogdonoff, Long Ouyang, Louis Feuvrier, Madelaine Boyd, Mamie Rheingold, Matt Jones, Michael Sharman, Miles Wang, Mingxuan Wang, Nick Cooper, Niko Felix, Nikunj Handa, Noel Bundick, Pedro Aguilar, Peter Faiman, Peter Hoeschele, Pranav Deshpande, Raul Puri, Raz Gaon, Reid Gustin, Robin Brown, Rob Honsby, Saachi Jain, Sandhini Agarwal, Scott Ethersmith, Scott Lessans, Shauna O’Brien, Spencer Papay, Steve Coffey, Tal Stramer, Tao Wang, Teddy Lee, Tejal Patwardhan, Thomas Degry, Tomo Hiratsuka, Troy Peterson, Wenda Zhou, William Butler, Wyatt Thompson, Yao Zhou, Yaodong Yu, Yi Cheng, Yinghai Lu, Younghoon Kim, Yu-Ann Wang Madan, Yushi Wang, Zhiqing Sun
Adam Brandon， Adam Koppel， Adele Li， Ahmed El-Kishky， Akila Welihinda， Alex Karpenko， Alex Nawar， Alex Tachard Passos， Amelia Liu， Andrei Gheorghe， Andrew Duberstein， Andrey Mishchenko， Angela Baek， Ankush Agarwal， Anting Shen， Antoni Baum， Ari Seff， Ashley Tyra， Behrooz Ghorbani， Bo Xu， Brandon McKinzie， Bryan Brandow， Carolina Paz， Cary Hudson， Chak Li， Chelsea Voss， Chen Shen， Chris Koch， 克里斯蒂安·吉布森、克里斯蒂娜·金、克里斯汀·麦克利维、克劳迪娅·费舍尔、科里·德卡罗、丹尼尔·雅各布维茨、丹尼尔·沃尔夫、大卫·凯尔克鲁德、大卫·李、埃赫桑·阿斯达尔、伊莱恩·金、埃米莉·古、埃里克·安东诺、埃里克·亨特、埃里克·华莱士、费利佩·托雷斯、福蒂斯·钱茨、弗雷迪·苏利特、詹巴蒂斯塔·帕拉斯坎多洛、哈迪·萨尔曼、包海明、王浩宇、亨利·阿斯佩格伦、郑亨元、伊恩·奥康奈尔、伊恩·索尔、伊莎贝拉·富尔福德、杰克·麦克尼尔、 詹姆斯·多诺万、杰米·基罗斯、杰森·艾、杰森·费多尔、杰森·魏、杰伊·迪克西特、杰弗里·韩、杰弗里·萨宾-松本、詹妮弗·格里菲斯-德尔加多、杰拉米·韩、耶利米·柯里尔、季林、韩佳佳、张佳明、翁佳怡、于洁琪、乔安妮·张、乔伊斯·鲁菲尔、陈凯、肖凯、凯文·巴顿、凯文·金、刘凯文、克里斯蒂安·格奥尔基耶夫、凯尔·米勒、拉玛·艾哈迈德、劳伦斯·福康内、伦纳德·博格多诺夫、 Long Ouyang， Louis Feuvrier， Madelaine Boyd， Mamie Rheingold， Matt Jones， Michael Sharman， Miles Wang， Mingxuan Wang， Nick Cooper， Niko Felix， Nikunj Handa， Noel Bundick， Pedro Aguilar， Peter Faiman， Peter Hoeschele， Pranav Deshpande， Raul Puri， Raz Gaon， Reid Gustin， Robin Brown， Rob Honsby， Saachi Jain， Sandhini Agarwal， Scott Ethersmith， Scott Lessans， Shauna O’Brien， Spencer Papay， Steve Coffey， Tal Stramer， Tao Wang， Teddy Lee， Tejal Patwardhan， Thomas Degry， Tomo Hiratsuka， Troy Peterson， Wenda 周， William Butler， Wyatt Thompson， Yao 周， Yaodong Yu， Yi Cheng， Yinghai Lu， Younghoon Kim， Yu-Ann Wang Madan， Yushi Wang， 孙志庆
Leadership 领导
Anna Makanju, Hannah Wong, Jerry Tworek, Liam Fedus, Mark Chen, Peter Welinder, Sam Altman
Anna Makanju、Hannah Wong、Jerry Tworek、Liam Fedus、Mark Chen、Peter Welinder、Sam Altman
