Today, we are launching Operator, an agent that can browse the web and complete tasks for you. It uses its own web browser to view web pages and interact by typing, clicking, and scrolling. Right now, Operator is in a research preview, so it has some limitations and will improve as we get feedback. Operator is one of our first agents and AIs that can handle tasks independently when given instructions.
Ask the Operator to automate browser tasks such as filling out forms, placing grocery orders, and generating memes. By working with familiar websites and tools, Operator streamlines daily workflows and creates new ways for businesses to engage customers.
We are starting with a small rollout to ensure a smooth launch. Operator is now available to Pro users in the US at operator.chatgpt.com. This research preview helps us learn and improve. As Operator develops, we plan to expand access to Plus, Team, and Enterprise users and add its features to ChatGPT. Now, let’s look at how Operator works.
How Operator Works
Operator runs on a new model called Computer Using Agent (CUA). CUA combines GPT-4O’s vision skills with advanced reasoning to work with graphical user interfaces, such as buttons, menus, and text fields you see on your screen.
Operator views your screen content through screenshots and interacts by performing mouse clicks and keyboard inputs within its browser. This enables it to execute web-based tasks without needing special API integrations.
If the operator encounters problems or makes a mistake, it can use its reasoning skills to resolve them. If it gets stuck and needs help, it gives control back to you, making sure the experience stays smooth and collaborative.
CUA is new and has limitations, but already sets records in Web Arena and Web Voyager, two key browser benchmarks. Read more about Operator’s research in our blog post. Now, let’s see how to use Operator.
How to Use
To start, just tell the operator what you want it to do, and it will take care of the rest. You can take control of the remote browser at any time. The operator is also trained to ask you to take over tasks that require a login, payment info, or CAPTCHA solving.
Customize the operator by adding instructions for its behavior across all sites or specific ones, such as setting flight preferences on booking.com. Save prompts for instant use on the homepage. Ideal for recurring tasks such as Instacart grocery restocks. Like browser tabs, initiate multiple operator sessions for parallel activities. For example, ordering a custom mug on Etsy while booking a campsite on Hipcamp.
Ecosystem and Users
Operator changes AI from a passive tool into an active part of the digital world. It helps users get things done faster and gives companies new ways to improve customer experiences and boost conversions. We’re working with companies like DoorDash, Instacart, OpenTable, Priceline, StubHub, Thumbtack, Uber, and others to ensure Operator meets real needs and complies with industry standards. We also see many ways operators can make certain tasks easier and more efficient, especially in the public sector. For example, we’re partnering with the City of Stockton to help people sign up for city services and programs more easily.
We’re releasing Operator to a small group to gather feedback and improve quickly. This approach balances new features, trust, and safety, ensuring Operator delivers value to users, creators, businesses, and public organizations.
Safety And Privacy
Protecting users is our top priority. Operator includes three safeguard layers to prevent abuse and keep users in control. Operator is designed for user control. It prompts for input at key moments.
- Takeover mode: when sensitive information, such as passwords or payment details, is required, the operator prompts you to take over. During this mode, the operator does not collect or record anything you type.
- User confirmations: Before actions such as ordering or emailing, the Operator asks for your approval.
- Operator declines sensitive tasks, such as banking or job applications.
- On sensitive sites, the operator requires close supervision to quickly address mistakes.
Operator provides simple controls for managing data privacy.
- Training opt-out: If you disable the improvement of the model for everyone in ChatGPT settings, the operator will not use your data for training models
- Transparent data management: The privacy section of operator settings lets you delete all browsing data and log out of all sites with one click. It also allows easy deletion of past conversations.
Protections are in place to prevent websites from attempting to mislead the operator with hidden prompts, harmful code, or phishing attempts.
Cautious navigation: operator recognizes and ignores prompt injections.
- A model monitors suspicious behavior and can pause tasks if it detects something wrong.
- Automated systems and specialists review threats and update safeguards quickly.
Operator is designed to refuse harmful requests and block unauthorized content. Moderation systems can warn users or revoke access if rules are repeatedly violated. Additionally, review steps have been implemented to detect and address misuse. Guidance is provided on using the operator in accordance with the usage policies.
No system is perfect, and Operator remains in a research preview. Ongoing improvements are informed by real-world feedback and thorough testing. Visit the Operator research blog’s safety section for more information.
Limitations
Operator is in an early research preview. It can complete many tasks but may make errors, especially with complex user interfaces such as slideshows or calendars. User feedback will inform improvements in accuracy, reliability, and safety.
What’s Next?
We plan to make CUA, the model behind Operator, available in the API soon. This will let developers build agents based on CUA. We’ll share a release timeline as we get more feedback from the research preview.
Enhanced capabilities: We’ll keep working to help Operator handle longer, more complex workflows. We’ll expand Operator to support 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, asynchronous task execution.
Source: Introducing Operator
