OpenAI adds a new AI agent to ChatGPT that can help users think and do tasks like scheduling appointments and creating presentations.
The ChatGPT agent combines many features from OpenAI's previous agents, including the ability to access Operator websites and aggregate information from dozens of websites into a concise research report from Deep Research.
Users can simply give natural language commands to use the new agent, which is currently rolling out to Pro, Plus, and Team subscribers. To activate it, they can select "agent mode" from the ChatGPT drop-down menu.
The launch of the new tool represents OpenAI's bold effort to turn ChatGPT into an actionable product and reduce the burden on users beyond just answering questions. Over the past few years, Silicon Valley companies like Google and Perplexity have introduced dozens of AI agents that promise to do just that. But the first AI agents struggled with complex tasks and weren’t as great as tech executives touted them to be.
OpenAI says its new ChatGPT agent performs much better than the company’s previous products. The tool “thinks” and “acts” using its own virtual machine, allowing it to handle complex requests. For example, a user can ask, “Look at my calendar and summarize my upcoming client meetings,” or “plan and buy ingredients to make a Japanese breakfast for four people.”
The ChatGPT agent scored 41.6% on the Humanity’s Last Exam (HLE), a 3,000-question test covering more than 100 categories, nearly double the scores of the o3 and o4-mini. On FrontierMath, one of the most difficult math tests, ChatGPT scored 27.4% when given access to tools, while the o4-mini model scored just 6.3%.
In a blog post, OpenAI acknowledged that ChatGPT's new features come with new risks. The company has implemented several safety measures, such as limiting the data the agent has access to and requiring user supervision for certain tasks, such as sending emails. The agent is also trained to reject "high-risk tasks" such as bank transfers.
Sam Altman, CEO of OpenAI, advised users to be cautious when allowing ChatGPT to access personal information. For example, granting calendar access to plan dinners may be reasonable, but the agent would not need calendar access to shop for clothes on the user's behalf.
(According to CNN, TechCrunch)