OpenAI has launched its inaugural artificial intelligence (AI) agent, named Operator, in a move announced on Thursday. As a research preview, this agent features a dedicated web browser and is designed to autonomously handle online tasks based on user prompts. The company revealed that Operator can assist with various activities such as booking tickets, making restaurant reservations, and purchasing products over the internet. At present, the agent is accessible exclusively to ChatGPT Pro subscribers within the United States, with plans for future availability across other subscription tiers.
OpenAI Unveils Operator AI Agent
In a recent live stream, OpenAI’s CEO, Sam Altman, took the opportunity to introduce the company’s first AI agent. He elaborated on the concept of AI agents, stating, “AI agents are AI systems that do work for you independently. You give them a task, and they go off and do it. We think it will be a big trend in AI.”
The Operator AI agent interface
Photo Credit: OpenAI
Operator is driven by the Computer-Using Agent (CUA), which integrates the vision capabilities of GPT-4o with enhanced reasoning, according to a blog post from OpenAI. The AI agent has undergone post-training through reinforcement learning and can interact with graphical user interfaces (GUIs) including buttons, menus, and text fields. Equipped with its dedicated browser, Operator can perform tasks discreetly in the background, allowing users to retain focus on their screens.
The AI agent accepts both textual and visual input. To execute tasks, the CUA processes screen pixel data and utilizes a virtual keyboard and mouse. OpenAI claims that the agent can handle multi-step tasks, troubleshoot errors, and adapt to unanticipated changes during its operations.
Potential Applications of the Operator AI Agent
Rowan Cheung, the founder of the AI newsletter The Rundown AI, had the chance to try out Operator early and shared its capabilities through several posts on X (formerly Twitter). One notable instance featured the agent planning a weekend trip by gathering information from Reddit, adhering to a specific budget and user interests. Remarkably, when access to Reddit was restricted, the agent pivoted by conducting a Bing search with “Reddit” as a search term to continue its task.
2. Planning a weekend trip based on hidden gems off Reddit, my budget and interests
Notice how at 0:06, ChatGPT Operator was blocked from Reddit but then decided to just do a Bing search with “Reddit” at the end
Very impressive decision-making pic.twitter.com/D5m3ouiiqt— Rowan Cheung (@rowancheung) January 23, 2025
In another scenario, Cheung tasked the Operator with identifying promising cryptocurrency tokens. The agent ran into a blockage at a CAPTCHA challenge labeled “Are you human?” and promptly alerted the user to confirm their identity. Once Cheung provided confirmation, the AI regained control and proceeded with the task.
Operator has been designed for seamless user interaction, allowing individuals to take over tasks at any point for edits or changes. Users can easily revert control back to the agent, ensuring they maintain authority over the AI at all times.
OpenAI has also announced partnerships with several companies such as DoorDash, eBay, Instacart, and Uber to comply with their terms of service while utilizing their platforms.
Safety Considerations and Mitigation Strategies for Operator
On the subject of safety, OpenAI reported that it has conducted extensive testing of the AI agent and has implemented measures to address three categories of safety concerns: misuse, model errors, and unpredictable risks.
To mitigate the potential for misuse, OpenAI has equipped the CUA model to reject harmful requests as well as those involving prohibited or regulated activities. This includes blocking access to gambling, adult content, and websites selling drugs or firearms. The company is also conducting both automated and manual reviews of user interactions for further safety assurance.
For concerns regarding model inaccuracies or hallucinations, the AI agent is trained to seek user confirmation prior to completing tasks with potential external impacts. The CUA refrains from assisting with banking transactions, and any attempts to access sensitive sites require vigilant user oversight.
Addressing frontier risks, which entail unexpected behavior from advanced AI models, OpenAI states that the CUA has been evaluated according to its Preparedness Framework, and the Operator System Card elaborates on its safety measures and ongoing enhancements.
At this time, Operator is exclusively available at operator.chatgpt.com for ChatGPT Pro subscribers located in the United States. The company has indicated plans to extend access to all ChatGPT clients in the future. Notably, a ChatGPT Pro subscription costs $200 (approximately Rs. 17,200) per month.
