From Developer Tool to Company-Wide Standard
What began as a coding assistant has quietly reshaped how every department inside OpenAI operates. The company's Codex platform — originally built for software development — now accounts for 99.8% of all output tokens generated internally across both Codex and ChatGPT combined, as of June 11, 2026. The figure comes from a research paper OpenAI released titled "The Shift to Agentic AI: Evidence from Codex," which draws on usage data across internal employees, external individual users, and external organizational accounts.
The takeaway is stark: conversational AI, for all its early momentum, has been effectively displaced inside OpenAI by an agentic model built around task delegation rather than dialogue.
Every Department Has Made the Shift
Engineering teams moved to Codex first, but they were not alone for long. By April 2026, non-technical departments including Legal, Finance, and Recruiting had crossed the threshold into majority Codex usage. Today, every department within OpenAI treats Codex as its primary AI interface.
The output data reveals just how dramatically workflows have changed. The median employee in a legal role generates 13 times more monthly output tokens than they did in November 2025. For researchers, that figure climbs past 50 times. These are not marginal productivity gains — they represent a fundamental restructuring of how knowledge work gets done.
The Agentic Shift: From Asking to Delegating
The core distinction the paper draws is between conversational AI and agentic AI. Conversational tools respond to prompts with advice, answers, or information. Agentic tools execute. As the paper puts it, users are asking Codex to do work — not simply to advise on it.
That shift in interaction model explains why token volumes have grown so dramatically. When a tool completes multi-step tasks autonomously rather than generating a single response, the computational footprint expands alongside the scope of what's possible.
Task Complexity Is Growing Alongside Usage
The data on task scope among external users tells the same story. As of the most recent measurement, 25.6% of sampled individual users had submitted at least one request estimated to require more than eight hours of equivalent human labor. In December 2025, that figure was just 2.1%. In under six months, the scale of what users are willing to hand off to an AI agent has grown more than tenfold.
Non-Developer Adoption Has Exploded
Among external users, the growth in non-developer adoption is particularly striking. Since August 2025, non-developer usage has risen 137 times among individual account users and 189 times among organizational account users. These are not incremental adoption curves — they reflect a category of workers who previously had little reason to use a coding-focused tool and who are now among its fastest-growing segments.
Organizational Accounts Outpace Individual Users
Not all external users are adopting at the same rate. Among organizational accounts, 17.3% of users have now adopted Codex. Among individual users, that figure is below 1%. The gap likely reflects the structural advantages enterprises offer: centralized access, guided onboarding, and institutional pressure to integrate new tools into existing workflows.
In terms of output token share among external users, organizational accounts attribute 63.3% of their AI output to Codex, while individual users remain at 16.5%. The enterprise adoption curve appears significantly steeper.
Codex Passes 5 Million Weekly Active Users
Weekly active users on Codex have now surpassed 5 million — a more than sixfold increase since February. That growth trajectory suggests the platform's expansion is not slowing as it scales, which stands in contrast to many enterprise tools that see adoption flatten after early enthusiasm.
OpenAI's Internal Data Comes With Caveats
The paper is careful to note what its internal data cannot tell us. OpenAI acknowledges that conditions inside the company are highly atypical: there are no usage restrictions, organizational buy-in is unusually high, and employees work in close proximity to the technology they are evaluating. Extrapolating from internal adoption rates to broader enterprise norms would be misleading.
That said, the directional signal — that agentic AI is expanding beyond its original developer audience and into every functional area of knowledge work — is consistent with what the external organizational data also shows. The specific numbers may not transfer, but the pattern does.