OpenAI says its agent products hit 10 million weekly users
OpenAI groups Codex and ChatGPT Work under an agent label and says they doubled in two weeks. The figures are self-reported and the definition of active is fuzzy.
- OpenAI says Codex and ChatGPT Work, the products it labels as agents, together hit about 10 million weekly users by mid-July, roughly double two weeks earlier.
- Per OpenAI, Codex alone crossed 7 million on July 13, up from about 1 million in February; non-developers are the fastest-growing cohort.
- Analysts flag the number as loose: OpenAI did not define active users, and promotional limit resets during launch likely inflated counts.
OpenAI is putting a big number on its agent push, and the number deserves an asterisk. The company disclosed that Codex and ChatGPT Work, the two products it groups under an agent label, together reached about 10 million weekly users by mid-July, roughly double the count two weeks earlier when ChatGPT Work launched on July 9. It is a steep curve for agentic products in actual work, and it is also a metric built on soft definitions.
The reported curve
Per figures attributed to OpenAI, the combined base passed 6 million on July 12 and reached 9 million within days, adding close to a million users a day after the release of GPT-5.6. Codex on its own crossed 7 million on July 13, up from roughly 1 million in February. OpenAI also says knowledge workers and non-developers are its fastest-growing cohort, now roughly one in five Codex users, which is the detail that matters most if you care about agents moving beyond engineering teams.
We covered the product that anchors the jump when it shipped, in our ChatGPT Work launch coverage. The two-week doubling lines up with that release window, which is exactly why the timing invites scrutiny.
Why the number is soft
Analysts have flagged the metric as loose, and the caveats are not minor. OpenAI did not specify whether active means daily or weekly, so the 10 million figure blends two things that are not the same. Promotional usage-limit resets during the launch almost certainly pulled in users who bumped against caps and came back, inflating the count during precisely the window OpenAI is measuring. And no specific enterprise customers were named, so there is no way to tie the growth to durable deployments rather than trial traffic.
None of that means the growth is fake. It means the shape of it is unverified. A number that doubles in two weeks around a launch and a promo reset is the kind of number that can also halve when the promo ends. You can read the reporting at Unite.ai and Crypto Briefing, both of which surface the same definitional gaps.
The signal under the noise
The cohort data is the part worth trusting more than the headline. If one in five Codex users really is a non-developer, that is the adoption pattern enterprises have been waiting to see, and it echoes the broad internal uptake we documented in Citi's 90 percent adoption figure. Agents earning their keep depends on people who do not write code reaching for them by default.
What this means for the agent economy
Ten million is a marketing number until OpenAI defines its terms, and buyers should treat it that way when it lands in a vendor deck. The durable question is not how many people opened an agent during a launch with the limits turned off. It is how many keep opening it in August, at full price, without a promo nudging them back. That retention curve, not the launch spike, is what tells you whether the agent economy is a business or a demo.