What Went Wrong With OpenAI’s Year of Agents?

A year ago, OpenAI CEO Sam Altman wrote in a blog post that he expected 2025 to bring the first AI agents that could join the workforce and “materially change the output of companies.” Altman, like so many other AI leaders at the time, was being overly optimistic.

As my colleagues Kevin and Sri have reported, OpenAI’s ChatGPT Agent, a product it launched in July to complete tasks on ChatGPT subscribers’ computers like building a financial model or buying ingredients for a dinner party, didn’t meet some internal goals, including being used by 10% of ChatGPT weekly active users.

At its peak, around the time of its launch, ChatGPT Agent reached four million paying ChatGPT weekly active users, said a person with knowledge of the effort. That was about 11% of the 35 million weekly active users paying for ChatGPT at that time (680 million people used the chatbot at least once a week at the time, mostly without paying anything).

A few months later, the agent product dropped below one million paid ChatGPT weekly active users, the person said. At the same time, the company cut its revenue expectations from selling agents by half, to $1.4 billion in 2025.

People didn’t quite know what to do with a general-purpose browser-using agent, said the person with knowledge of the effort. This reflects the bigger issue we’ve written about where ChatGPT users more broadly just don’t seem to understand the full range of things the chatbot can do, including that it can analyze a picture of a dying plant to give advice on how to revive it or a screenshot of a computer error to give advice on how to fix it.