Introducing Operator

A research preview of an agent that can use its own browser to perform tasks for you. Operator is currently available to Pro users in the U.S.

Operator is one of OpenAI’s first agents—AIs capable of working independently on your behalf. You give it a task, and it will carry out that task using its own browser. It can browse the web, look at pages, click buttons, scroll, and type just like a person would. Because it’s a research preview, Operator has certain limitations and will evolve over time based on user feedback.

Key Highlights

Availability and Purpose

• Immediate Availability: Pro users in the U.S. can try Operator now at operator.chatgpt.com.
• Scope of Tasks: Operator can fill out forms, order groceries, create memes, and handle a wide array of other repetitive browser tasks.
• Future Expansion: The plan is to eventually extend access to Plus, Team, and Enterprise customers, and to integrate Operator’s capabilities into ChatGPT.

How Operator Works

Operator leverages a new model called Computer-Using Agent (CUA), which combines GPT-4o’s vision features with enhanced reasoning. This allows Operator to:

• “See” through screenshots.
• “Interact” with a web interface just like a user—typing, clicking, and scrolling.
• Self-correct when it runs into obstacles or hand control back to the user when it’s stuck.

Early benchmarks show state-of-the-art results on WebArena and WebVoyager. More details on these evaluations and the research behind Operator are available in OpenAI’s research blog post.

Getting Started

Using Operator is straightforward:

1. Describe your Task: Tell Operator what you want done.
2. Monitor or Take Over: You can step in at any time to manually control the browser session, especially for private data entry (e.g., login credentials, payment details, CAPTCHAs).
3. Customize: Add custom instructions for sites or tasks. Save prompts for quick reuse.
4. Multi-Task: Open multiple conversations—like booking a campsite on one tab while ordering an enamel mug on Etsy in another.

Ecosystem & Users

Operator aims to turn AI from a passive helper into an active digital assistant. It can streamline tasks for users and offer businesses new ways to engage customers. Early collaborators include DoorDash, Instacart, OpenTable, Priceline, StubHub, Thumbtack, and Uber. Operator also holds potential for public-sector use cases, like simplifying enrollment in city services. The City of Stockton, for example, is exploring ways to leverage Operator for more accessible civic engagement:

“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.”
— Jamil Niazi, Director of Information Technology at City of Stockton

Early feedback from the private sector underscores its convenience:

“OpenAI’s Operator is a technological breakthrough that makes processes like ordering groceries incredibly easy.”
— Daniel Danker, Chief Product Officer at Instacart

Safety & Privacy

OpenAI has implemented three layers of safeguards to keep Operator usage safe and transparent:

1. User Control
   – Takeover Mode: Operator requests your intervention before entering sensitive data (e.g., passwords, payment info).
   – User Confirmations: Operator will ask for final sign-off on important actions.
   – Task Limitations: Certain sensitive tasks are restricted or declined.
   – Watch Mode: Particularly sensitive sites require direct user supervision.
2. Data Privacy Management
   – Training Opt-Out: Disabling “Improve the model for everyone” in ChatGPT also applies to Operator.
   – Transparent Data Management: You can clear your browsing data and log out of all sites with one click; past conversations can be deleted at any time.
3. Defenses Against Adversarial Sites
   – Cautious Navigation: Operator detects and ignores malicious prompt injections.
   – Monitoring: A specialized “monitor model” observes Operator’s actions for suspicious behavior.
   – Detection Pipeline: Automated and human reviews update safeguards against emerging threats.

Operator is trained to refuse harmful requests and block disallowed content. Repeated policy violations may lead to warnings or revoked access. As a research preview, Operator isn’t perfect, but OpenAI remains committed to continual improvement informed by real-world usage and ongoing testing.

Limitations

Operator is still in early development and may struggle with complex interfaces like slideshow creation or calendar management. Users can help shape future improvements by sharing feedback on accuracy, reliability, and safety.

What’s Next

• CUA in the API: Soon developers can build their own agents with the Computer-Using Agent model.
• Enhanced Capabilities: Operator will handle longer, more complex workflows.
• Wider Access: Once it’s proven safe and scalable, Operator will roll out to Plus, Team, and Enterprise users, and ultimately be integrated into ChatGPT.

Livestream Replay

If you missed the announcement livestream, a replay is available for viewing at operator.chatgpt.com. Operator’s evolution will continue to be guided by real-world user feedback and a commitment to balancing innovation with trust and safety.