Intel / article

How Codex went from cloud flop to ChatGPT's core

Source: every.to
Author: Dan Shipper
Date: Published Oct 15, 2026

Context

Dan Shipper, whose company Every tested OpenAI's models before launch, reconstructs how Codex went from a cloud product that failed to catch on to the foundation of the ChatGPT desktop app, drawing on months of interviews with Greg Brockman, Thibault Sottiaux, Alexander Embiricos, Mia Glaese and other team members (every.to). His central argument is that the team kept abandoning a working interface whenever the models improved enough to support a new one. Codex Cloud, launched in spring 2025, ran tasks on remote machines, but outside OpenAI developers had to recreate their setup in the cloud and could not interrupt a task mid-run, and the team later said the models were not yet smart enough. Under Sottiaux the team pivoted to running locally, rewrote the in Rust so it could sit behind any interface, and after GPT-5 launched, an editor extension for Codex became its most-used surface.

Rather than double down on that extension, the team built a desktop app for managing several agents at once and deliberately left out a code editor, so any change had to go through the agent. Shipper reports that the team itself was not using the app as its main tool until OpenAI's separate Codex model line improved in late 2025; by early January the whole team was. The app launched in February 2026, gained a browser and the ability to work across other Mac apps, and by June had more than 5 million weekly users, about one in five of them knowledge workers. On July 9, OpenAI made it the new ChatGPT desktop app, and Sottiaux now runs all of OpenAI's core products.

Shipper credits three things beyond the models: OpenAI running parallel internal bets and merging them once a direction proved out, researchers and product builders sitting on one team with their own model line insulated from ChatGPT, and the team using Codex to build Codex, which let a handful of people stand up a new product in weeks. He frames the sequence as a compressed version of the form-factor shifts Clayton Christensen described in The Innovator's Dilemma, and closes with Embiricos saying the next step is to move back to the cloud for asynchronous work, the vision that failed the first time.

Key takeaways · AI-distilled

Codex Cloud launched May 16, 2025 and drew little sustained outside use, per Dan Shipper: users had to recreate their laptop setup in the cloud, could not interrupt the agent, and each follow-up prompt started a fresh machine.
Thibault Sottiaux told Every there was no explicit decision to split off a Codex model line; it grew from internal agents trained on OpenAI's monorepo to speed up research infrastructure work.
After the cloud launch stalled, the team pivoted to local and rewrote the agent in Rust, separated from the terminal interface so the same core could power the CLI, the editor extension and later the desktop app.
The Codex desktop app shipped with no built-in code editor, so every change went through the agent; the team doubted that choice until GPT-5.1 Codex Max and GPT-5.2 Codex made the models good enough.
Per the article, Codex passed 2 million weekly users by March 19, 2026 and 5 million by June, about one in five a knowledge worker, before the Codex app became the ChatGPT desktop app on July 9.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source every.to