OpenAI reports Codex drives 99.8% of internal AI output tokens

• According to OpenAI's report thread, Codex now generates 99.8% of OpenAI's internal AI output tokens, up from under 10% a year earlier, while rohanpaul_ai's chart summary puts external shares far lower at 63.3% for organizations and 16.5% for individual users.
• rohanpaul_ai's non-developer growth chart says non-developer Codex adoption since August 2025 rose 189x in organizations and 137x for individuals, faster than developer growth outside OpenAI, while rohanpaul_ai's role breakout shows external usage is still concentrated in engineering and data roles.
• Task shape changed along with adoption: rohanpaul_ai's thread says 70.2% of sampled OpenAI users sent Codex work estimated above one human hour, 25.6% sent work above eight hours, and rohanpaul_ai's runtime chart shows the 99th percentile of internal users stacking more than 70 agent-hours per day.
• Reusable harnesses are part of the shift. rohanpaul_ai's skills summary says 96% of OpenAI weekly active users invoked at least one skill, versus 30% of organizational users and 26% of individual users, and emollick's note flags skills as one of the clearest signals in the dataset.
• The interesting gap is not whether Codex escaped engineering. It already did inside OpenAI, where gdb's department chart shows finance, recruiting, and legal near 90% Codex share, while rohanpaul_ai's manager-status chart shows external individual contributors using it more heavily than managers.

You can scan OpenAI's thread for the core charts, inspect gdb's department screenshot for how close legal and recruiting are to engineering inside OpenAI, and compare that to rohanpaul_ai's skills chart, where skill usage is almost universal internally but still early outside. thsottiaux's post also points to the February app release as the visible break in non-engineering adoption.

OpenAI's headline number is blunt: Codex now accounts for 99.8% of internal output tokens. OpenAI's report thread frames that as a company-wide shift, not an engineering-only one.

The external comparison is the useful part. According to rohanpaul_ai's chart summary, Codex reached 63.3% of output tokens for organizational users and 16.5% for individuals by June 2026, which makes OpenAI look less like a proxy for the market than an ahead-of-curve internal deployment.

That same inside-outside gap shows up by department. In gdb's department chart, engineering ends near 99% Codex share inside OpenAI, while finance reaches 91%, recruiting 89%, and legal 88%.

The fastest growth in the report is not from developers. rohanpaul_ai's non-developer growth chart says active non-developer usage since August 2025 rose 189x in organizations and 137x for individuals.

thsottiaux's thread points to February 2, the Codex app release, as the visible break in adoption outside engineering. The chart attachments in OpenAI's thread also mark the macOS app launch and free plus Go access in February, then the Windows app launch in March, before a knowledge-work app update in April.

The catch is that external adoption is still uneven. rohanpaul_ai's role breakout says engineering roles account for 27% average Codex share among organizational users, data roles 15%, while non-technical categories sit much lower even as their growth rates accelerate.

The report measures agents as delegated runtime, not just prompts. According to rohanpaul_ai's thread, 70.2% of sampled OpenAI users submitted work estimated above one hour of human effort, and 25.6% submitted work above eight hours.

Heavy users are also running multiple jobs at once. rohanpaul_ai's thread says 28.6% of OpenAI users managed five or more concurrent agents, while rohanpaul_ai's runtime chart shows the 99th percentile of internal users accumulating roughly 70 to 71 Codex-hours per day because overlapping turns are summed.

That is the most concrete evidence in the report that the work unit changed from single-turn assistance to parallel delegated execution.

Codex usage is not just more volume. It is also more structure. rohanpaul_ai's skills summary says 26.6% of all active users invoked at least one skill, but inside OpenAI that number reached 96% of weekly active users.

The skill mix matters:

• Any skill: 96% of OpenAI users, 30% of organizational users, 26% of individual users, per rohanpaul_ai's skills chart.
• Custom skills: 87% of OpenAI users, 19% of organizational users, 12% of individual users, per rohanpaul_ai's skills chart.
• Plugin skills: 92% of OpenAI users, 15% of organizational users, 14% of individual users, per rohanpaul_ai's skills chart.
• Preinstalled skills: 58% of OpenAI users, 11% of organizational users, 11% of individual users, per rohanpaul_ai's skills chart.

As emollick's note points out, the skills data is one of the clearest clues that firms are standardizing repeated agent workflows instead of treating Codex like a better chat window.

OpenAI's internal story is cross-functional. The outside world still looks more technical. rohanpaul_ai's role breakout says organizational usage is highest in engineering at 27% of output tokens and data at 15%.

Even so, the internal department growth curves show why OpenAI is pushing the broader office-work framing. rohanpaul_ai's department-growth chart shows median output tokens since November 2025 up 56x in research, 32x in customer support, 27x in engineering, and 13x in legal.

BlackHC's reaction and eliebakouch's note both land on the same oddity: research usage growth is the steepest line in the paper, while support and legal are rising fast enough to make this look less like a coding-tool adoption story than a general knowledge-work rollout.

One of the more specific external cuts in the report is manager status. rohanpaul_ai's manager-status chart says organizational individual contributors generated 12% of their output tokens on Codex, versus 5% for people managers.

That split matters because it adds a different adoption boundary than job function. OpenAI's charts suggest Codex is already spreading beyond engineers, but outside the company it is still landing first with the people doing the work directly, not the people coordinating it.