OpenAI is making a pitch it could not have made a year ago: that its coding agent belongs to people who do not code. In a new report shared first with Axios on June 2, the company says knowledge workers now account for roughly one-fifth of Codex users and are growing more than three times as fast as developers.
The numbers, and what they leave out
Codex has more than 4 million weekly active users, OpenAI says, up more than five times since the desktop app launched in February. That is the headline figure, and it is worth pausing on what "weekly active" actually counts. A user who opens the app once on a Tuesday counts the same as someone running it all day. The five-times growth is also measured against a February baseline that included a brand-new app with presumably modest adoption, so a big multiple is doing some heavy lifting.
The task-level breakdown is more interesting than the user count. OpenAI says data analysis is the fastest-growing knowledge-work task at 110% week over week, with research up 37% and what the company calls knowledge artifacts (reports, memos, contracts, PDFs, spreadsheets) up 36%. Week-over-week growth rates are volatile by nature, and OpenAI does not say which week. Still, the direction is clear enough.
One stat that needs less skepticism: more than 60% of users now run multiple Codex tasks at once at some point during the day, up from under half in mid-April. That is a behavior change, not a marketing number.
So is this a Cowork competitor yet?
Not quite. The whole story here is OpenAI chasing the lane Anthropic opened first. Anthropic released Claude Code in October 2025, watched holiday dabblers turn it into something viral, then shipped the more office-friendly Cowork app. OpenAI's Codex desktop app followed the month after Cowork.
Codex still feels like a developer tool wearing an office-worker costume. It connects to email, calendar, Slack and Teams, and a single click sets up a morning brief pulling your calendar and unread email. Useful. But the underlying thing is the same agent a backend engineer uses, which is exactly OpenAI's bet: that non-technical users will grow into the technical surface rather than need it sanded down. Anthropic bet the opposite, building a separate consumer interface. Both could be right on different timelines.
The expert who actually checked the work
The most honest voice in OpenAI's promotional moment is a Stanford professor it did not pay for. Andrew Hall, at the Graduate School of Business, used a coding agent to update a paper he had published five years earlier on universal vote by mail. The tool gathered new data, ran analyses, produced figures and drafted a new version with, in his words, not very much prompting.
"It did a lot right, which is kind of remarkable, but it made a number of errors."
That is Hall, after he hired a graduate student to audit the output by hand. The agent failed to collect all the data it needed and miscoded some of what it had. His verdict was that it still "very much needed an expert, PhD-level student to oversee it quite closely." Which is the part the growth charts skip over.
There is also a human cost surfacing among heavy users. Rootly CTO Quentin Rousseau told Axios that managing several agents at once gets more done but feels nothing like a satisfying day's work, comparing it to a gripping TV series that keeps you up at night rather than a marathon that tires you out. Andrej Karpathy, an OpenAI co-founder now at Anthropic, described being in a "state of AI psychosis" since December.
OpenAI wants to reframe Codex as something like an operating system for office work. Whether that holds depends less on the weekly active count and more on how many users discover, as Hall did, that the agent needs watching.
