OpenAI published a piece yesterday framing Codex as evidence that AI agents have crossed a threshold from developer tool to general workforce tool. The headline numbers: 97.9% of OpenAI employees now use Codex, up from roughly 40% in August 2025. Non-developer individual usage grew 137x. Non-developer organizational usage grew 189x. Long-task requests — those estimated to take eight or more hours of human work — grew tenfold. Every department at OpenAI, including Legal and Recruiting, apparently uses it as their primary AI tool.
The Register covered this with appropriate skepticism: every number here comes from OpenAI, a company with direct financial incentive to promote the product it is measuring. The figures are not independently audited. The methodology is opaque. The definition of “using Codex” could mean anything from completing a project to opening the interface once. The cautionary footnote applies.
With that said, there is something worth reading in the data that is harder to fake entirely.
The 137x non-developer individual growth figure is eyebrow-raising, but the more interesting signal is the resulting share: non-developers are now roughly 20% of Codex’s five million weekly users. That is a meaningful demographic shift even if the base was artificially low. Coding agents started as tools for software engineers because software engineers were the first to articulate requirements clearly enough for the early models to follow. The hypothesis has always been that once the models improved — and once the scaffolding for agentic task execution matured — the workflow benefits would generalize to anyone who does knowledge work with a computer. The Legal team’s 13x growth in token generation from November 2025 to June 2026 is one concrete piece of evidence that generalization is happening, even if 13x could mean “from essentially nothing.”
The 10x growth in long-task requests is separately interesting because it suggests users are changing what they ask for, not just asking more often. The shift from “write me a function” to “complete a task that would take a human half a day” represents a different trust relationship with the system. Whether the tasks are actually being completed reliably is a different question the data does not answer.
The epistemological problem runs deeper than just OpenAI having incentives. We are in a period where the only organizations that can measure AI adoption at meaningful scale are the companies building the AI. External researchers can run surveys, but they cannot see actual usage data. This means every “AI is transforming work” narrative is mediated by a financially interested party. That is not unique to AI — pharmaceutical companies have always reported their own clinical data — but it is worth naming as a structural feature of how the field currently communicates about itself.
The useful posture is: treat the absolute numbers with substantial skepticism, treat the direction as probably real, and watch for independent signals. One such signal: OpenAI claims Codex crossed five million weekly active users. The 20% non-developer share, if accurate, implies about one million non-developers using an AI coding agent weekly. That would be a genuine inflection. If enterprise adoption surveys from neutral parties start showing similar shifts in the next few months, the directional claim will have survived the credibility test. If they do not, the numbers were marketing.