For the first few months after OpenAI released Codex, its own employees mostly ignored it for daily work. Through August 2025, the average OpenAI worker ran less than 10% of their AI usage through Codex and stuck with ChatGPT for everything else. A year later, Codex is the primary AI tool in every department at the company, including Legal, Finance, and Recruiting, and the chatbot has become the sideshow.
The chatbot era ended inside OpenAI first
OpenAI's Economic Research team, led by visiting researcher Dave Holtz, mined the company's own Codex logs to trace the shift. The headline number: across the company, Codex now generates 99.8% of the AI output OpenAI produces in a given week. Engineers moved first, shifting the majority of their usage to Codex by December 2025. Legal, finance, and recruiting crossed over later, around April 2026, but their transitions were much faster. The average lawyer or recruiter at OpenAI now runs more than 85% of their work through Codex rather than ChatGPT.
The deeper change is what people ask for. Chatbot conversations tend to be short and self-contained. An agent can work on its own for minutes or hours, calling tools and iterating toward a result. So people stopped asking quick questions and started handing off projects.
Longer tasks, running in parallel
By May 2026, 80.6% of sampled individual users had given Codex at least one request estimated to equal more than 30 minutes of human work, and a quarter had handed it a task estimated at more than eight hours. At the extreme, OpenAI's heaviest internal users now generate more than 60 hours of agent work in a single day, spread across multiple agents running at once. People moved from asking for one answer at a time to orchestrating a small fleet of agents through the workday.
The fastest-growing users were never the target audience
The people adopting Codex fastest are the ones it was never built for: workers who don't code. Since August 2025, non-developer use rose 137-fold among individual users and 189-fold among organizations. Non-technical staff now lean on agents for automation, data cleanup, simple tooling, and analysis, and over a quarter of the Codex work done by people in business roles was actually engineering or coding they could not have pulled off before.
Ethan Mollick, who studies AI adoption at Wharton, called the data "a canary in the coal mine," adding: "The chatbot era is over, and agentic systems are coming to tasks beyond engineering."
The honest caveats
OpenAI is not a normal company. Its employees have free, easy access to frontier tools and every incentive to use them, so this is a preview of where things may go, while the average office today still looks very different. The study is also upfront that it measures usage, not results: it offers no estimate of whether all this delegation made anyone faster or their output better.
Codex isn't really a coding tool anymore. You talk to it like ChatGPT, but instead of just answering, it goes and does the whole task.
It's slowly becoming a super app, and it's interesting to see how OpenAI has adopted it internally. Something to consider for your own knowledge work too.
If you don't know where to start, I've made a couple of videos on getting the most out of it:
- The 7 Codex Plugins I Use Everyday
- Master 95% of Codex for Office Work (Full Course)
- Codex Just Quietly Changed How People Work Forever (Role-Specific Plugins)