Knowledge workers now make up roughly one-fifth of OpenAI's Codex users and are growing more than three times as fast as developers, according to a new OpenAI report shared first with Axios.
Why it matters: AI has made it easier to crank out documents, emails, decks and dashboards, and OpenAI is now betting agents can help workers make sense of them.
The big picture: Previous waves of workplace software encouraged workers to produce huge volumes of files and messages, but those "workplace artifacts" largely remain siloed inside different software programs.
- The report argues that Codex can round up the important context from all of those artifacts no matter where they are.
By the numbers: Codex now has more than 5 million weekly active users, up more than six times since OpenAI launched the desktop app in February, the company says.
- The fastest-growing tasks among knowledge workers are data analysis, up 110% week over week; research, up 37%; and knowledge artifacts — reports, memos, docs, contracts, multimedia assets, PDFs and spreadsheets — up 36%.
- More than 60% of users now run more than one Codex task at the same time at some point during the day, up from less than half in mid-April.
Case in point: Codex can connect to your email, calendar, documents, spreadsheets, design apps and messaging apps like Slack and Teams.
- It only takes one click to set up a daily automation that can send a morning brief that includes what's on your calendar, important unread emails, and anything else that Codex thinks needs your attention.
Catch up quick: Anthropic's Claude Code and Cowork were the first agentic tools to attract non-coders at scale.
- Anthropic released Claude Code as a command-line tool in February 2025, followed shortly thereafter by Codex.
- Anthropic was first with a consumer-facing web release in October 2025, and both released native desktop agent apps this year.
The other side: A growing number of power users say agentic tools are leaving them mentally fried, as they try to supervise several fast-moving AI workstreams at once.
- OpenAI co-founder Andrej Karpathy, now at Anthropic, told the "No Priors" podcast he had been in a "state of AI psychosis" since December, trying to figure out what was possible and "pushing it to the limit."
- Quentin Rousseau, CTO and co-founder of incident management platform Rootly, says using agents like Codex and Claude Code means getting more done. But, he says, the satisfaction that comes from a typical hard day's work is a lot different than the stress of managing agents.
- "It's kind of like the difference between running a marathon and watching, a really gripping TV series," he told Axios in March. "One tires you out and the other keeps you up all night."
Zoom in: Andrew Hall, a Stanford Graduate School of Business professor, tells Axios that he and his students use coding agents like Codex and Claude Code to help with boilerplate academic tasks, data collection, statistical analysis and running code to process data.
- Earlier this year Hall asked Claude Code to update a paper he'd published five years ago on universal vote by mail. "We figured papers like this should be updated over time, but no one ever does that," he says. The tool gathered new data, ran analyses, produced figures and tables and drafted a new paper, "with not very much prompting," Hall said.
- But when Hall hired a graduate student to audit the work manually, the agent's limits became clear. "It didn't do everything right," Hall said. "It did a lot right, which is kind of remarkable, but it made a number of errors."
- The tool failed to collect all the data it needed and didn't quite code all the data correctly, he said, meaning it "very much needed an expert, Ph.D.-level student to oversee it quite closely."
The bottom line: OpenAI is trying to reframe Codex from a tool for developers into something closer to an operating system for knowledge work.
Editor's note: This story has been corrected to show Codex's growth rate is 6x (not 5x), and to clarify the timeline of when various Anthropic and OpenAI tools were released.
