Industry

Use of OpenAI's Codex Agents Growing Outside Labs, Still Limited Among Consumers

A joint report by OpenAI, Columbia, Duke and the University of Pennsylvania finds increasing adoption of Codex, OpenAI's agentic coding and work platform, particularly among organizations and some individual power users.

Use of OpenAI's Codex Agents Growing Outside Labs, Still Limited Among Consumers

A new report from OpenAI, Columbia University, Duke University and the University of Pennsylvania shows that use of OpenAI's Codex — an agentic coding and work platform — is accelerating. The researchers divided Codex users into OpenAI employees, outside organizations and individual users, and compared the share of output tokens produced by Codex versus ChatGPT.

Key figures

  • Among OpenAI employees, 99.8% of output tokens were produced with Codex.
  • For outside organizations that use these tools, Codex accounted for 63% of output tokens.
  • For individual users, the share was 16.5%.

The report notes that OpenAI's internal usage offers a view of how users might adopt agents when barriers such as cost, access, training and organizational buy-in are largely removed.

Growth among organizations and individuals

Researchers found that among organizations active on both ChatGPT and Codex, Codex use was just above 0% in August 2025 and rose to about 17% by 2026. Although the absolute number of individual Codex users remains small, those individuals tend to use Codex intensively — a point first reported by Axios.

Individual-user sample details

The analysis for individual users is based on a 0.1% random sample of individuals who opted in to allow queries for training. Using model-based estimates:

  • 80.6% of sampled individual Codex users made at least one Codex request estimated to save more than 30 minutes of work for an "experienced human."
  • 70.2% made at least one request estimated to save over an hour of human work.
  • 25.6% delegated work estimated to take more than eight hours for a human to complete.

The report emphasizes these thresholds are model-estimated.

User profiles and timeline

While software work remains Codex's core use case, the fastest-growing user group is non-developers. The shift toward agentic work accelerated at the start of 2026, when general users began allowing Codex, OpenClaw and Anthropic's Claude Code to interact with desktops, manage calendars, read and write files, control web browsers, and run scripts.

How many consumers use Codex?

Most AI users are still primarily chatting with bots rather than running armies of agents. Among consumers on the Go, Free, Pro and Plus plans who were active on either ChatGPT or Codex in the last 28 days, fewer than 1% used Codex.

Reporter note

The article's journalist, who has covered cybersecurity for years and describes only basic early-2000s HTML skills, said they were initially reluctant to grant agents access to files, browser and apps. Over the past month, however, they began using Codex and Claude Code for many routine tasks: filling out expense reports, triaging email, booking hair appointments and reporting packages stolen from the apartment lobby — tasks that the reporter says have saved measurable time.

Expert perspective

Workplace culture expert Jessica Kriegel told Axios that agents are "reducing what I'd call the psychological cost of action." They make unfamiliar tasks feel more approachable, prompting earlier starts, more experimentation and less energy spent worrying about unknowns.

Conclusion

The joint report suggests Codex and similar agentic AI tools are being adopted more widely, particularly within organizations and by intensive individual users. Still, broad consumer adoption remains limited: most people continue to use AI for chat, and fewer than 1% of active consumer users engaged with Codex in the latest 28-day window.