OpenAI launches Operator, a web-browsing agent

Background

An AI agent is software that can carry out a task through multiple steps rather than only answering a single prompt. Operator was OpenAI's research preview of an agent that worked through a browser. It was not a general release: on 23 January 2025 it was available only to ChatGPT Pro users in the United States.^[1]

What OpenAI released

Operator used its own browser and could look at a webpage and interact with it by typing, clicking and scrolling.^[4] It was powered by the Computer-Using Agent model, which combined GPT-4o's vision with reasoning trained through reinforcement learning.^[2] In practice, the agent could fill in forms, place online orders and schedule appointments through the browser.^[3]

The described loop is visual and interactive: read what is on the page, decide the next step, then type, click or scroll. That moves the task beyond generating text on its own; the agent is acting on a web interface.^[4]

Key details and safeguards

Operator asked users to confirm before actions such as booking a hotel or sending an email.^[2] For login credentials or payment details, takeover mode handed control to the user; OpenAI said it did not collect or screenshot what was typed during that handoff.^[4] The agent was trained to decline sensitive tasks such as banking transactions. On sites such as email or financial services, it required users to watch its actions closely.^[4]

These safeguards show where the agent was allowed to act and where a person had to step in. Confirmation covered actions with consequences, while takeover covered secrets and payment. The sensitive-site rules put a boundary around banking and required supervision on email and financial services.^[4]

OpenAI described Operator as a research preview and said it planned to expand it to Plus, Team and Enterprise users and to integrate the capability into ChatGPT.^[4] At launch, the product was limited to Pro users in the US.^[1]

Impact

In July 2025, Operator's abilities were merged into ChatGPT agent.^[3] The merger placed the browser-operating capability inside ChatGPT agent.

Operator's launch is a marker in the move from chatbots that answer to agents that act. Its design also showed the guardrails that come with that move: confirmation prompts, a takeover mode for credentials and payment, a refusal policy for banking tasks, and supervision requirements on sensitive sites.^[4] Those are product decisions about how much autonomy to give a browser agent and where to keep a human in the loop.

Sources

^[1]: Wikipedia, "OpenAI". ^[2]: ForkLog, "OpenAI unveils AI agent Operator". ^[3]: Wikipedia, "ChatGPT". ^[4]: OpenAI, "Introducing Operator".