Inside OpenAI, the chatbot era is looking a bit old-fashioned. Employees are increasingly turning to Codex, the company’s agentic AI tool, for long-running, multi-step work instead of firing off one-off prompts to ChatGPT – and the shift is spreading well beyond software engineers.
That matters because internal habits at AI companies tend to leak outward. When a product becomes the default tool inside the lab that built it, everyone else tends to notice, including rivals, enterprise buyers, and the people trying to figure out whether AI is replacing tasks or just rearranging them. OpenAI also has a very obvious incentive to talk up the trend: agents are stickier, more expensive to use, and much easier to pitch as the future than a glorified question-and-answer box.
Codex is becoming the default tool inside OpenAI
OpenAI said active usage of agentic AI rose more than fivefold in the first half of 2026, with the fastest growth coming from outside its original software developer audience. The company added that by August 2025, the average employee was using less than 10% of their tokens on Codex, while now every department – including legal and HR – is using it as a primary AI tool at work.
The headline numbers are hard to ignore. OpenAI says 97.9% of its employees now use Codex, up from 40% of staff in August 2025. By contrast, just 17.3% of organizations outside OpenAI are using it today, and only 0.7% of private individuals do.
The longest tasks are getting the most attention
Codex is built to keep working for a while, which is exactly why it is moving beyond quick code suggestions. OpenAI said the share of private users who send at least one request for a task that would take an experienced human more than eight hours has risen almost tenfold since the start of the year. In plain English: people are increasingly asking AI to stay in the room until the job is done.
OpenAI also broke out growth by audience. Since August 2025, Codex use has jumped 137 times among non-software private users, 189 times among corporate users, and 12 times inside OpenAI itself. That is a neat reminder that the biggest growth often comes from the least obvious users, not the programmers everyone assumed would be first in line.
Legal and HR are joining the party
Most AI tools still get marketed as productivity helpers for drafting, searching, or summarizing. OpenAI’s internal data suggests the real shift is toward agents handling workflows that used to require repeated human prompting. The company said even nontechnical teams are now leaning on Codex, and in June 2026 the average OpenAI employee in a legal role generated 13 times more monthly tokens in Codex and ChatGPT than in November 2025.
That does not mean chatbots are dead. It does suggest the center of gravity is moving from asking AI questions to assigning it work. If that pattern holds outside OpenAI, the next fight in enterprise AI will not be about who writes the prettiest paragraph – it will be about which agent can survive the messiest workload without falling apart.