OpenAI's agents reach 10 million weekly users in less than two weeks

OpenAI claims that the agent part of ChatGPT has surpassed 10 million weekly users, a threshold reached in less than two weeks since the debut of ChatGPT Work, the system with which the company aims to transform the chatbot into a tool that completes a task rather than just describing it. The figure, initially reported by Bloomberg, combines the two products that OpenAI labels as agents: Codex, the assistant for code writing, and the newer ChatGPT Work, which performs multi-step tasks through user's apps and files.

10M!
New day, new usage reset for paid users of Codex and ChatGPT Work. Lands in the next hour. Enjoy.
pic.twitter.com/VUJ4S3vKDG
— Tibo (@thsottiaux)
July 21, 2026

It's important to clarify what that number really indicates: it refers to the weekly active users of both products combined, not the enterprise licenses, paying customers, or actual executed agents. It is also a company estimate, so it is not a certified metric, and OpenAI’s claims about usage have already raised skepticism in the past, including those self-declared that supported its internal research on adoption.

Even taking this into account, one cannot overlook how steep the curve is: according to counts that OpenAI executives published as each milestone fell, Codex surpassed 5 million weekly users by the end of May, reached 6 million on July 12, and around 8 million by mid-month. More than a million of these users now use Codex for tasks outside of software development, and the overall usage of agent tools has grown by approximately 2.5 times in the week following the release of GPT-5.6, the model powering ChatGPT Work. Sam Altman has also repeatedly reset usage limits as a push to increase adoption.

What does ChatGPT Work really do?

ChatGPT Work arrived on July 9 for Pro, Enterprise, and Edu subscribers, and quickly expanded to Plus and Business plans within days. Instead of just responding to a request and stopping, it breaks down a task into smaller steps and processes them, generating spreadsheets, presentations, documents, and small web applications, and staying on a project for hours. The agent connects to services like Slack and Google Drive, can schedule recurring tasks, and leverages OpenAI's Computer Use feature to operate on desktop programs on behalf of the user.

The pricing model mirrors that already implemented for Codex, based on consumption, and administrators of Enterprise and Edu plans can set spending caps, limit accessible tools, and track activity through a compliance interface. In parallel, the company announced a dedicated program for small businesses, featuring virtual assisted training, in-person sessions in the United States, introductory guides, and access to agents and partners designed for this customer segment.

Among the first customers mentioned at launch are Zapier, RingCentral, Virgin Atlantic, and Nvidia, with finance teams employing the tools to speed up month-end reporting and forecasting, automating data collection, reconciling figures, and constructing subsequent presentations. OpenAI also states that nearly all its staff now use agent products daily.

Momentum numbers, revenues still to be proven

The timing places OpenAI in a crowded race, where Google has just launched Gemini Spark, a constantly active assistant that takes instructions through Gmail, while Anthropic has pushed its agents for coding and office work. All are chasing the same promise: that an AI system can be tasked with completing an activity rather than just sketching it out, and they compete for the same corporate budgets and the limited pool of IT managers willing to assign real work to software that still requires supervision.

What the figure flaunted by OpenAI cannot clearly indicate is whether the usage will endure over time or convert into revenues at the level that OpenAI needs as it approaches its much-discussed IPO. The company is experimenting with other avenues to monetize ChatGPT, including advertising, and a weekly user count says little about how many of the 10 million sit on paid plans or how often they return once the novelty of a new model fades. For now, the number should be interpreted as a simple momentum indicator, self-reported and taken mid-journey, rather than proof of a market already won.