OpenAI's internal ChatGPT is essentially abandoned, with 99.8% of AI output now handled by Codex.

June 26 news: OpenAI's published economics paper "The Shift to Agentic AI: Evidence from Codex" reveals the transformation of knowledge work patterns by tracking data from the agent platform Codex. The basic unit of work is shifting from single conversations to asynchronously delegating longer and more complex tasks to AI.

Within OpenAI, agents have completely taken over daily work, and the traditional chat interface ChatGPT is nearly obsolete. As of June 2026, 99.8% of tokens generated by internal employees each week come from Codex. Among them, engineers' average Codex token usage accounts for 99%, while non-technical employees in legal and recruiting also exceed 85%. Deep usage has led employees to run AI like coordinating teams, with nearly 28.6% of employees managing more than 5 agents simultaneously, and the heaviest users averaging over 71 hours of daily cumulative runtime.

Compared to OpenAI's extreme state, external users are also accelerating their shift to agents. Sampling of external individual users shows that the proportion of human work tasks estimated to take over 8 hours submitted within half a year rose from 2.1% to 25.6%. Since August 2025, the number of non-development weekly active users among external individuals and organizations surged 137 times and 189 times respectively. Additionally, the usage rate of Skills, which center on reusable workflows, among all active Codex users jumped from 5.4% to 26.6% within three months.