ChatGPT Agent Launches: AI Takes on Complex Tasks Independently

OpenAI has just unveiled the ChatGPT Agent — a new agent mode that integrates existing features from Operator and DeepResearch with entirely new technologies. According to the company’s president, Sam Altman, the launch of this mode signifies a shift from «AI discussing» to «AI executing.»

The Agent is accessible via the ChatGPT interface (both web version and mobile app), where users can assign it complex tasks. One demonstration involved planning a trip for a friend’s wedding, during which the Agent was tasked with selecting various outfit options that matched the wedding’s style, finding a hotel for the necessary dates, and making a reservation. Once given the assignment and after clarifying the details, the Agent activates a virtual computer that provides access to various tools, including a web browser and connections to Google Drive, OneDrive, and others.

Throughout its operation, the Agent can ask additional questions and subsequently returns with a report detailing its work — in this case, an itinerary, outfit recommendations, and hotel suggestions. Once decisions are made, users can instruct the Agent to order the selected clothing and book the hotels.

In another example, the Agent was tasked with locating benchmark results on Google Drive and converting them into a presentation. A noteworthy aspect of this demonstration was the Agent’s ability to «see»: after creating an initial version, it assessed the design and made adjustments to the final layout. Interestingly, the agent mode outperformed OpenAI’s language models in benchmarks — for instance, it scored 42% in Humanity’s Last Exam, placing it on par with Grok 4 and significantly ahead of ChatGPT o3.

The agent mode will be available today or tomorrow for all paid ChatGPT subscribers, with Plus users receiving the ability to complete 40 tasks per month.

P.S. You can support me by subscribing to the channel “escaped neural network,” where I discuss AI from a creative perspective.