Every says Codex and ChatGPT for Work have 40 million weekly users combined

OpenAI’s Codex and ChatGPT for Work together have 40 million weekly active users, according to an October 15 retrospective by Every’s Dan Shipper. The figure puts a new scale on a product story that began with a struggling cloud coding agent and led to Codex’s integration into the ChatGPT desktop app.

Shipper interviewed six people involved in Codex’s research, engineering, product and design, including Greg Brockman, Alexander Embiricos and Amelia Glaese. His account traces more than a change of interface: the team built dedicated coding models and a desktop app designed to manage several agents at once.

Coding gave agents a way to check their work

The technical foundation was a reasoning model: a system trained to work through a problem before returning an answer. Shipper describes how that approach let a model consider different paths, backtrack and revise, rather than work out its answer only while writing the response a user sees.

Programming offered a useful feedback loop. An agent could edit code, run it and inspect the result, then attempt another fix. That gave it a way to test changes without asking a person to judge every step. But turning that loop into a useful product also meant deciding how long users should wait and how closely they should supervise.

The editor was a success, not the destination

By summer 2025, the cloud version of Codex had failed to find product-market fit, Shipper writes. The team was shifting toward tools that ran on developers’ own computers, both in a terminal—a text-based command interface—and inside the editors where programmers write and debug code.

That route attracted users. After GPT-5 launched, OpenAI shipped a Codex extension for editors including Visual Studio Code and Cursor. Codex usage grew more than tenfold by October 2025, and the extension overtook the terminal as its most-used interface, according to the retrospective.

Yet Shipper’s own team wanted something different from close, step-by-step collaboration. In his account of testing GPT-5 before launch, Codex felt too quick to stop. His colleagues were already using Anthropic’s Claude Code to hand over whole tasks and wait for completed work. That was Every’s experience, not a controlled comparison of the products.

The Codex team did not stop at the editor. It trained a separate line of coding models from those powering ChatGPT and built the desktop app around managing multiple agents. The product direction moved beyond helping someone through one coding exchange toward overseeing several pieces of delegated work.

40 millionCombined weekly active users

Every’s October 15 figure covers Codex and ChatGPT for Work together, rather than Codex alone.

The audience was already moving beyond developers

OpenAI’s June 2 announcement supplies an earlier adoption marker: Codex had more than 5 million weekly active users, up more than sixfold since the desktop app’s February launch. Developers remained its largest user group, but the company said knowledge workers represented about 20 percent of users.

Those workers were using Codex to produce reports, spreadsheets, presentations and contracts, OpenAI said. Research, data analysis and workflow automation were other uses. The company also described users running multiple tasks in parallel, applying the same delegation model to work outside software development.

By July, Codex had merged into the flagship ChatGPT desktop app, Shipper writes. That history helps explain the combined audience now being reported, but it also sets a boundary on the comparison: June’s figure measured Codex alone. The new total does not show how much either product grew individually or how the audience divides between them.