99.8% of all output tokens generated inside OpenAI now come from Codex agents, not ChatGPT. That number alone should stop any engineer who still thinks agents are a gimmick.
OpenAI's economic research paper—published today—drops the internal data that proves the shift from chatbots to delegated, long-horizon agents is not theoretical. By June 2026, the average OpenAI worker generates 85% of their output tokens through Codex. The remaining 0.2% of ChatGPT tokens are a rounding error in a company that builds the models.
Non-developers overtook engineers in Codex adoption speed
Engineering moved first, as expected. By December 2025, the average OpenAI engineer had switched majority of their AI usage to Codex. Today, that number sits at 99%. But the real story is Legal, Finance, and Recruiting—they crossed over around April 2026, and their transition was faster. Now every department, including non-technical ones, uses Codex as their primary AI tool.
Non-developer individual users grew 137x since August 2025. Organizational non-developer users grew 189x. Inside OpenAI, non-developer adoption jumped 12x. That's not a slow creep; that's a land rush. Lawyers at OpenAI now generate 85% of their output tokens with Codex, running automation scripts and data transformations.
Agents work eight-hour tasks while you sleep
80.6% of sampled individual users made at least one Codex request that would take a human over 30 minutes. 70.2% exceeded one hour. 25.6% made a request exceeding eight hours of human labor—a full work day. The longest horizon tasks grew the fastest from a low base.
Among OpenAI's daily active users, the 99th percentile now regularly generates over 60 hours of Codex agent turns per day—distributed across parallel agents. Users aren't asking for one answer; they're orchestrating fleets of agents that run for hours. This is the unit of work changing from a single chat turn to a delegated project.
Codex usage deepened exponentially across departments
Since November 2025, median output tokens per active user rose 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal. These are not linear improvements. Agents compound their own usage as models get stronger and parallelization becomes standard.
The paper's key insight: agents don't just do existing tasks faster—they enable workers to take on tasks outside their job description. Non-technical users debug code, transform datasets, and build tools they never would have touched six months ago. The boundary between developer and non-developer is dissolving.
Expect every organization that deploys agents to see the same pattern: chatbots become legacy overnight, and the real growth comes from the non-developers who suddenly gain leverage on technical work.
Source: How agents are transforming work
Domain: openai.com
Comments load interactively on the live page.