How Codex Happened
Every
OpenAI’s coding bet looked shaky last summer, right before GPT-5 shipped. Then Codex went from timid sidebar tool to the core of ChatGPT.
Based on reporting by Every — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
I can point to the night the mood shifted. It was summer 2025, on the eve of GPT-5, in OpenAI’s Mission Bay office. The launch had that big-company electricity, but the thing we’d been testing for a week in upstate New York felt oddly restrained. GPT-5 was smart and fast inside Codex, yet it seemed to stop itself before doing the full job.
That mattered because the coding world was moving. At Every, we had started seeing a new way to work with models: not step-by-step help, but full delegation. With Anthropic’s Claude Code, some of us would hand over a task and come back later to a finished feature, or even a whole app. Compared with that, GPT-5 in Codex felt boxed in, like it wanted permission from the sandbox before it dared to be useful.
OpenAI’s team knew the product had to move. Codex had already missed on its first cloud version, so after GPT-5 launched the team shipped a sidebar extension for VS Code and Cursor. That kept the product near the editor, where programmers already lived, and it worked: by last October, usage had grown more than 10-fold, and the editor extension had become Codex’s main surface.
But the bigger change was deeper than placement. OpenAI started training and improving a separate line of coding models, built a desktop app for juggling multiple agents, and then pushed Codex into the center of the company’s products. By June 2026 it had more than 5 million weekly active users and had spread beyond programming into other knowledge work. By July, it was folded into the flagship ChatGPT desktop app, and Thibault Sottiaux had been put in charge of OpenAI’s core products and platforms. Today, Codex and ChatGPT for Work together have 40 million weekly active users. The story is not that OpenAI found a better autocomplete. It found a way to make waiting look like progress.
The clue was hiding in plain sight: “Thinking.” Once models could spend time on a problem, coding became the obvious place to cash that out. Software gives the model a loop — write, run, check, revise — and that loop is much more useful than a polite assistant that quits early.
My take — AI-written commentary, not fact-checked reporting
OpenAI didn’t just build a coding tool; it tripped over the fact that the future of software is less typing and more waiting for a machine to come back with something usable. That is either progress or a very expensive way to discover that programmers wanted a junior engineer, not a clever autocomplete. The real competition here is not who writes the prettiest line of code, but who can tolerate giving the model the room to think.
Read more about this at: Every
