OpenAI finds 97.9% of employees now use Codex agents, not ChatGPT chats

OpenAI is making a quiet but consequential claim: 97.9% of its employees now use Codex, up from around 40% in August 2025. And the point is not that people are “using AI more.” It is that they are switching the way they use it, from chat to agents that can tackle multi-step work that runs for hours.

This is happening inside OpenAI and is also spreading outside it, according to the company paper behind the announcement. OpenAI reports active usage growing rapidly, with the number of active users growing more than fivefold in the first half of 2026, and the most rapid increase occurring outside the initial audience of software developers. For executives trying to model costs, compliance exposure, and ROI, this matters because longer-running tasks are a different economic animal than one-off questions. When work takes hours, tokens get burned differently, and governance has to keep up.

So what exactly is “agentic AI usage” in this context? OpenAI describes employees shifting away from one-off ChatGPT prompts toward using Codex agents to take on multi-step tasks that take long periods of time. In other words, instead of asking an AI for the first step, employees are increasingly asking it to carry the work forward through a longer sequence. The company also says it sees similar behavior among external organizations, and even some individual usage, although individuals are still comparatively low.

The adoption numbers OpenAI provides underline that the change is not just for developers. Within OpenAI, it says 97.9% of employees use Codex. Externally, organizations are at 17.3% presently, while individual usage is about 0.7%. Within the company, OpenAI adds that “every department,” including non-technical departments like Legal and Recruiting, uses Codex as their primary AI tool for work. That detail is what turns this from a feature update into a workforce shift. When legal and recruiting teams become “primary users” of an engineering-adjacent tool, AI deployment stops being a developer productivity project and starts resembling an internal operating system.

OpenAI also tries to connect this shift to economics, and there is a specific angle executives will care about: token consumption and billing. Longer running tasks consume more tokens, and to the extent those can be billed, that should help “diminish hundreds of billions in debt obligations,” per the source text. The logic is straightforward. Agentic work tends to last longer. Longer sessions tend to produce more compute and therefore more tokens. If the market monetizes that usage, revenue per user can rise relative to costs. That is why OpenAI insists its findings matter beyond employee behavior, potentially reaching labor researchers and policymakers as well.

OpenAI does not immediately answer the obvious question of incentives, at least not in this report. It did not respond to a request to clarify whether it incentivizes or encourages employees to use its tools. The source notes possibilities such as internal communiques, token allocations, token use leaderboards, or tying tool usage to performance metrics, but it does not claim any of those are true. Still, OpenAI’s own blog post data is specific: “Through August 2025, the average OpenAI worker spent less than 10 percent of their tokens on Codex.” Now, according to OpenAI, every department uses it as the primary AI tool.

The behavioral evidence goes further than adoption rates. OpenAI’s paper, titled “The Shift to Agentic AI: Evidence from Codex” (PDF), says that since the start of the year, the share of individual Codex users who submit at least one request for a task estimated to require more than eight hours for an experienced human to complete has increased nearly tenfold. That is a major clue about what is changing operationally: tasks that used to be broken down into a series of human-guided steps are increasingly being packaged into longer AI-run workflows.

But OpenAI also flags an important nuance: comparing “human time” estimates to model time is only part of the picture if workflows are not fully automated. Code can be generated faster, sometimes at a higher rate, but code verification and deployment may still require humans. The second-order effect for enterprises is that agentic AI can compress the creation phase while leaving the quality, review, and integration phases intact, at least for a while. That means ROI can rise quickly in some workflows, while risk and review burden can rise in others, especially if agents generate more complex outputs that require more scrutiny.

Finally, OpenAI connects agentic adoption to role-based productivity in a way regulators and labor watchers will notice. Since August 2025, OpenAI says non-developer usage of Codex has risen 137x for individuals, 189x for organizational users, and 12x within OpenAI. It concedes technical usage remains dominant, but argues that adoption by non-devs shows broader knowledge workers taking on coding or technical tasks like automation, data transformation, and data analysis. It also provides a concrete internal datapoint: in June 2026, the median OpenAI employee in a legal role generated 13 times more monthly output tokens across Codex and ChatGPT than they did in November 2025.

Layer that onto the regulatory environment. The source notes that the number of US federal lawsuits filed against OpenAI and associated entities grew about 11 percent, from 35 to 39, between the last six months of 2025 and the first six months of 2026. The report suggests this timing, alongside the legal team’s 13x token surge, looks like OpenAI’s legal team is making the case for AI productivity benefits. Even if you strip away any narrative framing, the practical implication is clear: when agents become standard tools in legal workflows, compliance and dispute handling can shift from “reviewing AI outputs” to “managing AI-driven processes.” For executives, that means AI governance cannot be a generic policy document. It has to reflect how agents actually behave in long-running, cross-department work.