theregister.com web signal

OpenAI: 97.9% of Employees Use Codex, External Adoption at 0.7%

openai coding tools agents enterprise-ai coding-tools ai-agents

TL;DR

  • OpenAI employee Codex adoption reached 97.9%, up from roughly 40% in August 2025, per the company's own research paper.
  • Non-developer Codex usage rose 137 times for individual users and 189 times for organizational users since August 2025.
  • External individual user adoption stands at just 0.7%, far below the near-universal uptake OpenAI reports internally.

When OpenAI published its research paper "The Shift to Agentic AI: Evidence from Codex," the headline figure was striking: 97.9 percent of the company's employees now use Codex, up from roughly 40 percent in August 2025. As The Register reports, the shift is not confined to developers. Non-developer Codex usage reportedly rose 137 times for individual users and 189 times for organizational users since August 2025, with the company's own legal team generating 13 times more monthly output tokens in June 2026 than in November 2025.

The figure that deserves as much attention as the internal adoption rate is the external one. Individual external user adoption sits at 0.7 percent, with organizational external usage at 17.3 percent. Active agentic AI users across the platform grew more than fivefold in the first half of 2026 by OpenAI's account, but the gap between near-universal internal use and minimal individual external uptake is a real constraint on how far the "agentic era" framing travels beyond OpenAI's own offices.

There is a financial dimension the paper raises explicitly and that is worth sitting with. OpenAI notes that longer-running multi-step tasks consume more tokens, which the company describes as a potential revenue opportunity. That framing does not make the adoption numbers wrong, but it does mean the incentive to report them favorably is present. The reporting also notes that OpenAI did not disclose whether internal structures incentivize or require employees to use Codex, which would matter considerably for interpreting that 97.9 percent figure.

For teams considering where agentic AI currently delivers, the legal department data is the most concrete example the paper offers: measurably more output from non-technical users over a sustained period. What the paper does not give you is anything on output quality, error rates, or the amount of human review the generated work still requires before it is usable.