OpenAI Agents Blog Reveals 85% of Output Tokens Now Flow Through Codex
View original source →On June 25, 2026, OpenAI published 'How Agents Are Transforming Work' — a data-rich blog post providing the most detailed public data yet on how agentic AI has transformed an organization's internal operations.
Key findings:
• 85% of all output tokens for the average OpenAI employee now flow through Codex agents
• By June 2026, median Codex usage at OpenAI Research is 56 times higher than in November 2025
• Users at the 99th percentile regularly generate more than 60 hours of Codex agent turns per day — the equivalent of more than two full work months of agent labor daily from a single high-intensity user
• Legal, Finance, and Recruiting teams — not just engineers — moved to majority Codex agent use in April 2026
• Non-technical departments, after an initial adoption curve of 3-4 months, achieve comparable or higher per-capita agent utilization rates as engineering teams
This data destroys the assumption that agentic AI adoption requires a technical team as the primary user base. The highest early adoption rates came from teams with the highest repetitive task loads (Legal contract review, Recruiting candidate screening, Finance report generation) rather than teams with the highest technical sophistication.
Why It Matters: This empirically proves that organization-wide agent rollouts are feasible on 6-12 month timelines when implementation accounts for non-technical user onboarding. The 'AI is just for developers' narrative is demonstrably wrong at the organizational level.