OpenAI’s coding agent Codex flopped on release in May 2025 and now sits at the center of the company’s products, and the account of why is more instructive than the usage number attached to it. Dan Shipper, who runs the publication Every, reported the history in a long feature published on 15 October, drawing on interviews with the people who rebuilt it.
The first version, Codex Cloud, ran on a remote machine. You described a task, it copied your code to an isolated cloud computer, edited files, ran tests, and returned a summary and a proposed change. Shipper describes three problems. A follow-up request spun up a brand-new remote machine each time. A user could not step in to redirect the work. And before any of it ran, setup meant copying credentials, installed packages, permissions and preferences from a personal computer to the cloud, which could eat hours. Inside OpenAI it was “very successful internally,” Thibault Sottiaux told him, because the company’s standardized setups made the hours of configuration disappear. Their own tooling had masked the burden they were placing on outside users.
Embiricos also blames timing: “It was too early because the models weren’t smart enough.” Researcher Mia Glaese concedes the team chased a future vision instead of meeting developers “on their laptop.” The piece says outside interest was real but sustained use was thin, and gives no usage figures for Cloud.
The turnaround has four named parts. Sottiaux took over the team as engineering manager and, in his words, “pivoted to local,” which put the agent on the same machine the programmer was already working on. He then spent the first months rewriting the agent underneath, in the Rust programming language, and separating it from the terminal screen so the same engine could later sit in an editor plug-in or an app. Meanwhile Michelle Pokrass’s post-training group watched programmers work, found where the model broke on messy real code, and turned those failures into training tasks, which produced a careful model that paused to ask permission.
The fourth decision is the least obvious. The editor plug-in was growing fast (usage rose more than tenfold by last October, per the piece), and most companies would have doubled down. The team instead built a desktop app with no code editor at all. To fix anything, you had to ask the agent. Designer Andrew Ambrosino wrote that this was the launch scope he was “most skeptical about,” and early on none of the four builders used it as their main tool. It changed once newer Codex models arrived in late 2025. The public release came on 2 February 2026.
The 40 million figure. The article states that “Codex and ChatGPT for Work combined have 40 million weekly active users.” That is not a Codex number. The piece does not say who produced it, does not define “active,” and does not say how much of it is Codex. The figures it attributes to OpenAI are smaller: 2 million weekly Codex users by 19 March, and more than 5 million by June, roughly one in five of them knowledge workers rather than programmers. Those are the numbers to use for Codex itself.
Whose account this is. Shipper interviewed Embiricos, Sottiaux, Glaese, OpenAI president Greg Brockman, Ambrosino and others, so the turnaround is told by the people who carried it out, and Altman’s “most loved internal product we’ve ever had” is a quote about internal use. Shipper also says Every tested GPT-5 under early access from OpenAI, and he switched much of his own work to Codex. He is a satisfied customer as well as a reporter. His own August 2025 worry, that OpenAI was building a tool for a dying workflow, is one the piece shows to have been off.
My reading, not the article’s finding: the Codex story is a warning about internal enthusiasm. The cloud product passed the test of the people who built it and failed the test of strangers. The desktop app passed both tests, though the piece describes outside adoption mostly through OpenAI’s own figures. A product team’s own adoption proves little until someone with a different laptop, and no OpenAI setup scripts, has tried it.
For teams building agent products, the useful audit is the first-hour cost: if a new user needs hours of setup before the agent does anything, the people who built the setup cannot see that cost.
Reported by Every on 15 October 2026
