At OpenAI, the typical worker now has AI agents running on their behalf for about two and a half hours every single day.

That's the median.

The heaviest users? They have roughly 71 hours of agent work happening inside a single day. That number only makes sense once you realize they aren't running one agent. They're running a whole crowd of them at the same time, the way a manager watches over a team that never sleeps, never eats, and never goes home.

This comes from a new OpenAI study of how people actually use Codex, its agentic coding and work tool. And I think it's the clearest evidence yet that the way we use AI is changing shape.

The study looked at three groups: everyday users, companies using Codex, and OpenAI's own employees. Around the same time, OpenAI also previewed a new model called GPT-5.6 Sol. It's reportedly so good at cybersecurity that the U.S. government asked OpenAI to delay its public release.