Codex Is No Longer Just a Tool for Programmers

OpenAI is betting that AI agents will be able not only to help create documents, emails, presentations and dashboards, but also to make sense of their contents. Earlier generations of office software made it easier to produce vast numbers of files and messages, but those workplace materials still often remain scattered across different programs.

The report argues that Codex can gather important context from such materials regardless of where they are stored.

According to OpenAI, Codex now has more than 4 million weekly active users—more than five times as many as after the desktop app launched in February.

Among knowledge workers, the fastest-growing use of Codex is data analysis, up 110% week over week. Next come research tasks, growing by 37%, and work with so-called “knowledge artifacts”—reports, memos, documents, contracts, multimedia files, PDFs and spreadsheets. Growth in that category reached 36%.

More than 60% of users now run multiple Codex tasks at the same time at least once during the day. In mid-April, fewer than half did so.

Codex can connect to email, calendars, documents, spreadsheets, design apps and messaging platforms such as Slack and Teams. A daily automation, for example, can be set up with one click: it will send a morning briefing with the schedule, important unread emails and other matters that Codex believes require attention.

The first agentic tools to attract a mass audience beyond programmers were Claude Code and Cowork from Anthropic.

Anthropic released Claude Code in October 2025. During the winter holidays, users began actively experimenting with it, and early in the new year Claude Code went viral. Later, Claude itself wrote Cowork, an app more oriented toward office tasks. OpenAI released the Codex desktop app the following month.

But the trend also has a downside. A growing number of active users say agentic tools overload them, because they have to supervise several fast-moving AI processes at once.

OpenAI co-founder Andrej Karpathy, now working at Anthropic, said on the No Priors podcast that since December he has been in a “state of AI psychosis,” trying to understand what is possible and “pushing it to the limit.”

Quentin Rousseau, chief technology officer and co-founder of the incident-management platform Rootly, says agents such as Codex and Claude Code really do make it possible to do more. But, he says, the satisfaction of an ordinary hard workday is very different from the stress of managing agents.

“It’s roughly the difference between running a marathon and watching a very gripping TV series,” he said in March. “One exhausts you, while the other keeps you up all night.”

Andrew Hall, a professor at Stanford Graduate School of Business, said he and his students use Codex and Claude Code for routine academic tasks: preparing boilerplate code, collecting data, conducting statistical analysis and running programs to process large datasets.

Earlier this year, Hall asked Claude Code to update a paper on universal mail-in voting that had been published five years earlier. “We decided that papers like this should be updated over time, but nobody ever does it,” he said. The tool gathered new data, ran the analysis, prepared charts and tables and wrote a draft of a new paper—with “not very much prompting,” according to Hall.

But when Hall hired a graduate student to manually check the work, the agent’s limitations became clear. “It didn’t get everything right,” Hall said. “It got a lot right, which is itself pretty remarkable, but it made a number of mistakes.”

According to him, the tool did not collect all the necessary data and did not code the entire dataset quite correctly. Its work therefore “very much needed an expert—a PhD-level graduate student—carefully supervising it.”

OpenAI, meanwhile, is trying to rethink Codex: not as a tool only for developers, but as something closer to an operating system for knowledge work.