OpenAI has unveiled its latest innovation, ChatGPT Agent, AI feature designed to perform complex and multi-step tasks on behalf of users. This new tool can manage schedules, plan events, shop online, create presentations, and even analyze data—all by using its own “virtual computer.” The feature is now rolling out to Pro, Plus, and Team subscribers, with Enterprise and Education users expected to gain access later this summer.
Announcing the launch OpenAI said, “ChatGPT can now do work for you using its own computer, handling complex tasks from start to finish.” Companies like Google, Meta, and Klarna are also investing heavily in AI agents—systems that do much more than chatbots by taking real action on a user’s behalf.
In earlier versions of ChatGPT, the Agent can browse the web, run code, interact with APIs, and connect with tools like Gmail and Google Calendar. It can complete tasks based entirely on user instructions, working in the background while keeping users informed of its actions.
To start using it, subscribers can enable the feature by selecting “agent mode” from ChatGPT’s tools dropdown. Pro users will receive 400 messages per month, while Plus and Team users will get 40 messages monthly.
OpenAI shared examples of what the new agent can do in a demo video. The AI was seen planning a romantic evening by checking Google Calendar for the user’s availability, searching for restaurant reservations on OpenTable, and even booking a table—all while letting the user make changes during the process. In another example, it analyzed the popularity of rival toy brands Labubus and Beanie Babies and created a detailed slide deck.
The ChatGPT Agent operates using a combination of advanced tools, including a text browser, a visual browser, and a terminal. These allow it to run code, navigate websites, and interact with connected apps. The system functions in a sandboxed virtual machine on OpenAI’s servers, meaning it does not have direct access to the user’s device or private files.
An OpenAI engineer described it during a livestream: “ChatGPT carries out these tasks using its own virtual computer, fluidly shifting between reasoning and action to handle complex workflows from start to finish, all based on your instructions.”
This ensures users maintain control. Before taking actions with real-world consequences, like making purchases, the agent asks for user approval. There is also a “Watch Mode” feature for tasks that require active oversight.
According to OpenAI, ChatGPT Agent achieves impressive results in benchmark tests. On Humanity’s Last Exam, which challenges AI with expert-level questions, it scored 41.6 percent, nearly double the performance of OpenAI’s earlier o3 model.
In math, it achieved 27.4 percent accuracy on the FrontierMath benchmark, outperforming previous systems. The company also claims it performed better than humans in some data science tasks.
However, OpenAI admits the system is not perfect. In one simulated scenario, the agent struggled to complete a complex operation in a mock online retailer network. While it performed initial research steps well, it failed to chain together more advanced actions. This shows that the AI is still learning and may stumble with unfamiliar tasks.
With ChatGPT Agent’s ability to take direct actions online, safety and privacy are top concerns. OpenAI said it has built safeguards to prevent prompt injection attacks, where malicious websites might try to trick the AI into unsafe behavior.
The system requires user confirmation for suspicious actions and is trained to refuse high-risk tasks like bank transfers. It also has overseer models monitoring its activity to spot potential issues.
For privacy, since the agent operates on OpenAI’s servers, it does not access local data. Users can delete all browsing data and log out of active sessions with a single click. During “takeover mode,” OpenAI says it does not collect or store any user-entered data, including passwords.
Pro users can start using ChatGPT Agent today, while Plus and Team users will gain access over the next few days. Enterprise and Education plans are expected to get it in the coming weeks. Access for users in the European Economic Area and Switzerland is still pending.
The company also confirmed that its earlier Operator research preview site will remain active for a few more weeks before being shut down. OpenAI noted, “This is not just a chatbot anymore. It’s a system that can act on your behalf, making your digital life easier, faster, and smarter.”
