OpenAI Data Shows Codex Agents Are Transforming Knowledge Work
A new OpenAI economic research paper puts hard numbers behind a shift many teams are starting to feel: agentic AI work has moved well beyond the developer's desk. Among a sample of individual Codex users, 25.6% submitted at least one request by May 2026 that OpenAI estimates would take a person more than eight hours to complete. The basic unit of knowledge work is changing, from short chatbot replies to long jobs handed off to an agent.
Titled "The Shift to Agentic AI: Evidence from Codex," the paper documents four usage trends across individual subscribers, enterprise organizations, and OpenAI's own staff over the past year.
The Growing Shift Toward Long-Horizon AI Tasks
The eight-hour figure sits at the far end of a wider pattern. By May 2026, 80.6% of sampled individual users had made at least one Codex request OpenAI estimated would take a person more than 30 minutes. A further 70.2% crossed the one-hour mark.
Those estimates come from an LLM-as-judge system reading Codex transcripts, and OpenAI says they should be treated as directional rather than exact. Even so, the paper reports that nearly a quarter of all Codex requests now map to tasks of more than an hour of human work. A year earlier, that would have been a stretch for a tool built to write code.
At the most intensive end, users at the 99th percentile generated more than 60 hours of Codex agent turns per day by June 2026, spread across multiple agents running in parallel.
Legal, Finance, and Recruiting Crossed Over in April 2026
OpenAI's own workforce gives the clearest before-and-after picture. Engineering moved first, and the average engineer there now generates 99% of their output tokens through Codex rather than ChatGPT. Legal, finance, and recruiting reached their tipping point around April 2026, and the average lawyer or recruiter now directs more than 85% of output tokens to Codex.
Company-wide, Codex accounts for 99.8% of weekly output tokens generated internally, a share OpenAI attributes partly to Codex users burning through far more tokens per session than non-users.
Business teams are also stepping outside their job titles. More than a quarter of the Codex work done by finance and business-operations staff fell into the engineering or coding category, work that used to need dedicated technical support. A finance analyst can now build their own tooling. A recruiter can run a data transformation.
Non-Developer Codex Adoption Has Grown 189 Times Since August 2025
The fastest growth is happening away from engineering. Non-developer users on organizational accounts have grown 189-fold since August 2025. Individual users outside engineering multiplied 137 times over the same window, while non-developer adoption inside OpenAI, starting from a higher base, reached 12 times its August level.
Not all of that adoption looks alike. Most non-developers are using Codex for some form of long-horizon knowledge work, the paper notes, rather than turning into full-time software developers.
For businesses redesigning workflows and researchers tracking the future of work, OpenAI argues the trend is not unique to its own culture. The pattern reflects what happens when capable agentic tools become broadly and cheaply accessible: people drift from short chatbot exchanges toward longer, harder, more cross-functional delegation. You can read the full study in OpenAI's announcement, How agents are transforming work.