OpenAI Introduces Operator: An AI Agent for Online Tasks

With Operator, OpenAI has laid the groundwork for a new era of AI interaction—one where models don’t just generate information but take action.

2 mins read
OpenAI CEO Sam Altman Visits "Making Money With Charles Payne" at Fox Business Network Studios on December 04, 2024 in New York City

OpenAI has launched Operator, its first AI agent capable of performing online tasks within a web browser, marking a significant milestone in AI-driven automation. Powered by the new Computer-Using Agent (CUA) model, Operator can handle a variety of tasks, such as booking concert tickets, filling grocery orders, or making restaurant reservations. Currently, the tool is exclusively available to ChatGPT Pro users in the U.S. through OpenAI’s $200-a-month premium plan, with plans for broader availability in the future.

The release of Operator places OpenAI in direct competition with other leading AI tools like Anthropic’s Computer Use and Google DeepMind’s Mariner. According to OpenAI, Operator outperforms these rivals across several industry benchmarks, positioning it as a frontrunner in the emerging race to dominate agent-based AI.

“This is a natural evolution of what large language models can do,” said Yash Kumar, a researcher at OpenAI. “Moving from generating text to executing tasks unlocks new possibilities for productivity and problem-solving.”

How Operator Works

Operator interacts with websites just as a human would, navigating graphical user interfaces like buttons, text boxes, and menus. Instead of relying on specialized APIs, which are limited to specific applications, Operator scans the screen, performs an action, then reevaluates its progress step by step. This approach enables the model to work with almost any website.

For example, during a live demonstration, Operator used OpenTable to book a restaurant reservation, StubHub to find concert tickets, and Instacart to fill a grocery list based on a photo of handwritten items—all without human intervention.

Because it operates in the cloud via a remote browser on OpenAI servers, Operator can manage multiple tasks simultaneously. OpenAI is also collaborating with several companies, including DoorDash, Uber, Instacart, OpenTable, and StubHub, to streamline its functionality for specific use cases.

Performance and Safety

OpenAI has extensively tested CUA, the model behind Operator, against benchmarks designed to evaluate an agent’s ability to perform tasks. On OSWorld, which measures task efficiency in areas like file management and image editing, CUA scored 38.1%, outperforming Anthropic’s 22.0%. On WebVoyager, a browser-specific benchmark, it achieved 87%, edging out Google DeepMind’s Mariner at 83.5% and far surpassing Anthropic’s Computer Use at 56%.

Safety was a key focus during development. OpenAI employed red-teaming methods to explore potential misuse, such as bypassing ethical safeguards or interacting with maliciously designed websites. If Operator encounters uncertainty, it halts and asks users for clarification before proceeding with actions that could have external consequences.

A Glimpse of the Future

Although Operator is still in its early stages and prone to occasional errors, its capabilities represent a promising leap forward in the utility of AI agents. Kumar, who uses Operator in his personal life, highlighted its convenience for routine tasks: “I can delegate things like booking a date night or managing my grocery list in seconds, saving time and avoiding forgetfulness.”

Despite its potential, Operator is not without controversy. OpenAI’s partnerships with companies like OpenTable and Instacart raise questions about how deeply integrated such AI tools will become in commercial ecosystems.

The New Frontier for AI

The debut of Operator underscores a broader trend among top AI firms, with OpenAI, Anthropic, and Google DeepMind all racing to develop advanced agent-based models. According to Ali Farhadi, CEO of the Allen Institute for AI, this shift is transformative. “It’s impactful enough that people will use it, and constrained enough for the technology to work well in its current state,” he said.

With Operator, OpenAI has laid the groundwork for a new era of AI interaction—one where models don’t just generate information but take action. For now, it’s a tool for the privileged few, but its implications could soon reshape how we interact with technology.

Sri Lanka Guardian

The Sri Lanka Guardian is an online web portal founded in August 2007 by a group of concerned Sri Lankan citizens including journalists, activists, academics and retired civil servants. We are independent and non-profit. Email: editor@slguardian.org

Leave a Reply

Your email address will not be published.

Latest from Blog