OpenAI today introduced Operator, a new AI agent that can perform tasks on the web independently. Users simply give it instructions, and it completes the task without human intervention.
“AI agents are AI systems that can do work for you independently. You give them a task, and they go off and do it,” said OpenAI chief Sam Altman.
Simply put, the Operator can navigate websites, fill out forms, and make purchases—all by interacting with the web like a human. Unlike traditional automation tools that rely on APIs, Operator processes information visually, moving a virtual mouse and typing into a browser.
“Before, if you wanted your model to buy stuff from Instacart, you’d need to figure out if Instacart had an API… Now, this is just using screenshots, no API, nothing,” said OpenAI’s Yash Kumar during the demo.
Initially available for Pro users ($200 monthly ChatGPT Pro plan) in the US, Operator will expand to other regions, though European availability will take longer due to regulatory challenges. Altman, however, said that the company would make the tech “better, cheaper, and more widely available soon.”
Also, Operator will be released in OpenAI’s API “in the next few weeks.”
“2025 is the year of agents,” said OpenAI’s Greg Brockman. “Operator — research preview of an agent that can use its own browser to perform tasks for you.”
The Challenges: Not Perfect Yet
While Operator is impressive, it’s not flawless.
During the live demo, it made mistakes, such as selecting the wrong location for a restaurant booking. OpenAI admits that errors—sometimes embarrassing ones—are part of the early research phase. “Operator is an early research preview. It will do a lot of cool things. It also makes mistakes, sometimes embarrassing ones,” Kumar noted.
Safety is another concern. AI navigating the web independently could fall for scams, make incorrect purchases, or misinterpret user intent. To address this, Operator includes safeguards such as human confirmations and fraud detection.
“What if the website is misaligned? Maybe it’s fraudulent or asks Operator to wire money… We’ve developed our model to avoid those instructions, but we also have a separate layer—like an antivirus—that monitors suspicious activity,” explained OpenAI’s Reiichiro Nakano.
How is it different from Anthropic’s Computer Use?
OpenAI isn’t the only company working on AI agents. Anthropic recently launched ‘Computer Use,’ a feature in Claude 3.5 Sonnet, which allows AI to navigate computers like humans—using a cursor, clicking buttons, and typing text.
Both ‘Operator’ and ‘Computer Use’ share a similar goal: enabling AI to interact with digital systems as a human would. However, the key difference is accessibility. While ‘Computer Use’ is primarily available through API integrations for developers, Operator is directly accessible to consumers through ChatGPT. “We really want to put it in people’s hands,” OpenAI emphasised.
Performance also varies. In OSWorld, a test that evaluates AI’s ability to use computers, OpenAI’s COUA model scored 38.1%, while Claude 3.5 Sonnet scored 14.9%. This suggests Operator may be more reliable for real-world tasks, though both systems are still in development.
Meanwhile, Perplexity, known for its AI-powered search engine, has taken a different approach with Perplexity Assistant. Unlike Operator, which focuses on web navigation, Perplexity Assistant is designed for mobile devices. As Perplexity describes it, “Perplexity Assistant uses reasoning, search, and apps to help with daily tasks.”
A key advantage of Perplexity Assistant is its deep integration with smartphone workflows. It can search the web, book appointments, and even use a phone’s camera to identify objects. Unlike Operator, it maintains context across tasks, allowing users to research restaurants and book reservations in one seamless flow.
However, Perplexity has struggled with reliability in past features. As one report noted, “Perplexity launched half-baked products in the past. For instance, our testing found that Perplexity’s shopping feature… tended to be slow and error-prone.” The company acknowledges these issues, with Srinivas from Perplexity stating, “Some Perplexity Assistant actions [might] not always work.”
While Perplexity Assistant competes more with Google Assistant and Siri, Operator is positioned as a tool that could disrupt traditional web-based tasks.
What about Microsoft’s Copilot Vision? Microsoft has also entered the AI assistant space with Copilot Vision, but its approach is distinct from Operator and Perplexity Assistant. Instead of performing tasks, Copilot Vision enhances browsing by reading pages, summarising content, and offering insights in real-time.
As Microsoft describes it, “Copilot can now understand the full context of what you’re doing online. When you choose to enable Copilot Vision, it sees the page you’re on, it reads along with you, and you can talk through the problem you’re facing together.”
Unlike Operator, Copilot Vision does not take independent actions like making bookings or purchases. It simply provides guidance while browsing, much like an intelligent companion. Privacy is also a key focus—Copilot Vision is opt-in, and all browsing data is deleted once a session ends.
