Usage of Codex, OpenAI's agentic coding and work platform, is accelerating, according to a new report from researchers at OpenAI, Columbia University, Duke University, and the University of Pennsylvania. The study categorizes users into OpenAI employees, outside organizations, and individuals, measuring token output against ChatGPT. This signals a shift from AI-powered chat and web search toward delegated, autonomous work.
The findings suggest that the frontier AI labs' long-standing promise of effective digital assistants may be materializing. OpenAI's own Codex use is intended to model how users might adopt agents once barriers like cost, training, and access are minimized. Among active ChatGPT and Codex users at outside organizations, adoption has climbed from near zero in August to approximately 17%.
Within OpenAI, 99.8% of employee output tokens were generated using Codex, compared to 63% for organizations and just 16.5% for individuals. While the absolute number of individual Codex users remains small, those who adopt it tend to use it heavily, the report notes. These metrics illustrate a stark divide in agent tool adoption across different user groups.
For organizations, the rapid uptick in Codex usage suggests that AI agents are moving from experimental tools to mainstream productivity aids. This could reshape workflows in software development and other knowledge work sectors. However, the low individual adoption rate highlights that consumer markets remain nascent for agent-based AI products.
The study's authors caution that the sample of individual users is limited, making broad generalizations difficult. Wider adoption may require further reductions in cost and complexity before agents achieve the ubiquity of current chat-based AI tools.