AI Agents Are Now the Primary Work Tool at OpenAI
AI agents transforming work is no longer a prediction. At OpenAI, it is already the reality, and the company has the numbers to prove it.
OpenAI published a new economic research paper on June 25 documenting how its agentic coding tool, Codex, has moved from a developer utility to the dominant AI tool across every department in the company, including Legal, Finance, and Recruiting. Codex now accounts for 99.8% of weekly output tokens generated internally at OpenAI, up from less than 10% just under a year ago.
What makes this shift significant is not just adoption numbers. It is what people are actually doing with the tool. Nearly a quarter of all Codex requests involve tasks that would take a human more than an hour to complete. By May 2026, over 80% of individual users had made at least one request estimated to represent more than 30 minutes of human work, and more than 25% had submitted tasks equivalent to over eight hours of effort.
The paper highlights a particularly striking pattern around non-developer growth. Since August 2025, non-developer individual users of Codex grew 137 times over. Among organizational users outside OpenAI, the figure was 189 times. This means people in roles like marketing, finance, and operations are now using an agentic AI tool, originally built for engineers, to handle work that once required technical specialists. More than one-fourth of Codex output from business function workers fell into the engineering or coding category.
Inside OpenAI, the heaviest users at the 99th percentile were generating more than 60 hours of Codex agent activity in a single day by June 2026, running multiple parallel agents simultaneously. The research team saw the biggest internal growth, with median usage rising 56 times between November 2025 and June 2026. Customer Support rose 32 times and Engineering rose 27 times over the same period.
OpenAI frames this internal data as a preview of where the broader workforce is headed. As agentic tools become more capable and accessible, the company argues that workers will increasingly move from short chatbot interactions toward delegating longer, more complex tasks to AI agents operating independently over extended periods. The research is positioned as early evidence of what happens when capable agentic tools meet low-friction access at scale.
The findings arrive alongside related OpenAI developments, including a partnership with Samsung Electronics to bring ChatGPT and Codex to its employees and the unveiling of a new LLM-optimized inference chip with Broadcom, both pointing to an accelerating push to embed agentic AI deeper into enterprise workflows.
