OpenAI reports Codex became its primary work tool across every department

Why it moved the index

The study provides direct operational evidence that an AI developer shifted engineering and non-technical departments toward parallel agent labor, with Codex producing more than 85 percent of average-worker output tokens and 99.8 percent company-wide. More than 70 percent of sampled individual users delegated at least one task estimated above one hour of human work, and 25.6 percent delegated one above eight hours. This materially strengthens evidence that agents are absorbing implementation-heavy knowledge work and accelerating frontier-lab productivity. It does not demonstrate net job losses, and task horizons are model-estimated rather than directly timed, which limits magnitude and confidence.