OpenAI Says Codex and ChatGPT Work Hit 10 Million Users

OpenAI says 10 million people are now using Codex and ChatGPT Work, the two products it files under a single “agent” label — roughly double the count from two weeks earlier, when it launched ChatGPT Work. Codex lead Tibo Sottiaux posted the milestone on Tuesday morning with another usage reset attached, and the company shared the figure with Bloomberg ahead of a planned announcement the same day, July 21, 2026, framing it as evidence that its push into autonomous, task-completing software is gaining ground on rival Anthropic.

The headline number is real and first-party. What it measures is harder to pin down — and for a company that discloses little about how its products are actually used, that gap is the more interesting part of the story.

What the 10 million counts

OpenAI reports the figure as a combined tally of active users across Codex, its coding agent, and ChatGPT Work, the broader task agent it launched on July 9, 2026. It does not break out how many people use each, and it has not said whether “active” means daily, weekly, or a point-in-time snapshot.

Part of that ambiguity is a product decision. Since the July 9 launch, Codex and ChatGPT Work live inside one redesigned ChatGPT desktop app and draw from a single usage pool, so the line between a “Codex user” and a “ChatGPT Work user” is blurred by design. OpenAI also began retiring its standalone Atlas browser the same day, folding more of its agent surface into one app. The cleaner disclosures came earlier: in early June 2026, the company put Codex alone at more than 5 million weekly active users — a standalone, defined metric. The 10 million figure is a looser one.

Why the curve steepened

The accelerant was GPT-5.6, the model family OpenAI made generally available the same day it shipped ChatGPT Work. Its flagship tier, Sol, is tuned for coding, knowledge work, and long-running tasks, and demand climbed fast once it opened to paying users. Codex lead Tibo Sottiaux tracked the run in public: the combined base passed 6 million on July 12, 2026, then climbed through 7 and 8 million and past 9 million over the following days — close to a million new users a day.

Some of that steepness is engineered. As each milestone landed, OpenAI reset usage limits for all users and kept the five-hour rate cap off for Plus, Business, and Pro subscribers, nudging people toward heavier sessions. Reset mechanics like these pull lapsed users back into an active window, which lifts any point-in-time “active” count. It is a reason to read the trajectory as a demand signal during a promotional window rather than a clean measure of durable adoption.

The composition of the growth is the part OpenAI most wants noticed. The company says its fastest-growing cohort is non-developers — the roughly one in five Codex users, as of early June, who are knowledge workers rather than engineers. That shift is the entire rationale for ChatGPT Work: taking an agent built to write code and repositioning it as software that drafts documents, builds spreadsheets, and assembles presentations for people who never open a terminal.

That reach carries a familiar tradeoff. An agent that acts across a user’s files, email, and apps for hours at a stretch raises the same oversight questions OpenAI has flagged in its own long-horizon model testing — how much autonomy to grant, and where a human still needs to approve what the system does.

The race with Anthropic

The disclosure is timed to a competition OpenAI is not clearly leading. ChatGPT Work goes directly at Anthropic’s Claude Cowork, and GPT-5.6 Sol against the Claude model line — both aimed at the enterprise buyers who pay the most for AI, and both arriving as OpenAI prepares for a public listing that makes momentum figures useful. Within hours of OpenAI’s earlier milestone posts, Anthropic extended promotional pricing on its Fable 5 model and raised Claude Code’s weekly usage limits by half, a mirror-image retention play. OpenAI has also leaned on GPT-5.6’s safety testing as a selling point to enterprise and government buyers wary of agent autonomy.

For anyone weighing which agent to standardize on, the 10 million figure demonstrates traction, not superiority. It says a large number of people opened these products during a heavily promoted launch; it says nothing, by itself, about how many stay once the free resets stop, or how the finished work holds up. The benchmark numbers OpenAI cites for GPT-5.6 are vendor-reported, and independent evaluations of these agents on real tasks remain thin. The usage curve is the loudest number OpenAI has to offer right now — which is exactly why it is worth reading with the definition attached.