ChatGPT can now shop for you, schedule meetings, and analyze presentations with new agent feature
After launching Deep Research and Operator agents earlier this year, OpenAI has introduced a new AI agent called the ‘ChatGPT agent’, which integrates more deeply with the company’s chatbot of the same name. The ChatGPT agent is a general-purpose agentic system that can take actions on a user’s behalf.
With the new ChatGPT agent, the chatbot can handle tasks such as accessing the user’s calendar, shopping on their behalf, creating spreadsheets, and browsing the web to gather information. It can also automate repetitive tasks, such as converting screenshots or dashboards into presentations, planning and booking offsites, updating spreadsheets with new data, and more.
OpenAI says the ChatGPT agent brings the best of both worlds from the Operator and Agent tools, allowing it to transition naturally from a simple conversation to executing actions within the same chat.
ChatGPT can access your Gmail and GitHub
Amid rumours of OpenAI launching its own AI browser following the release of Perplexity’s Comet, the company has instead added advanced browsing capabilities to the ChatGPT agent. OpenAI describes this as a “visual browser that interacts with the web through a graphical user interface, a text-based browser for simpler reasoning-based web queries, a terminal, and direct API access.”
The ChatGPT agent can also utilise ChatGPT connectors to access websites like Gmail and GitHub, enabling it to find relevant information around a prompt and use it in its responses. Users will have the option to log into different websites at any point by taking over the browser.
The new agent manages tasks using its own virtual computer, which OpenAI says “preserves the context necessary for the task, even when multiple tools are used.”
The ChatGPT agent is designed to be more flexible and interactive than previous models. Users can interrupt it at any point to clarify instructions, steer it towards a desired outcome, or change the task entirely. The agent can also seek additional details from the user if needed to complete a task.
ChatGPT agent outperforms Copilot on spreadsheets
On benchmarks, OpenAI says the ChatGPT agent scores higher than the o3 reasoning model and Deep Research agent on Humanity’s Last Exam. On FrontierMath, a benchmark featuring novel, unpublished problems, it reportedly achieves better accuracy than the o3 and o4-mini models.
Notably, on SpreadsheetBench—a benchmark evaluating a model’s ability to edit real-world data in spreadsheets—OpenAI claims that the ChatGPT agent not only outperforms its own previous models but also Microsoft’s Copilot.
ChatGPT agent’s risk profile is higher
OpenAI acknowledges that while it has built numerous safeguards and warnings into the new agent, its overall risk profile remains higher than previous models.
OpenAI CEO Sam Altman noted on X (formerly Twitter) that the ChatGPT agent could be tricked by “bad actors” into revealing users’ private information.
“We are going to warn users heavily and give users the freedom to take actions carefully if they want to,” he added.
How to use the ChatGPT agent
The ChatGPT agent can be accessed by selecting ‘agent mode’ within the ChatGPT app or website. The feature is currently available only to OpenAI’s paying members, including Pro, Plus, and Team users.
