OpenAI Codex Users Surge 5X as Non-Developers Drive Adoption

Among external organizational users, OpenAI's Codex adoption is growing across both managers and non-managers, but individual contributors are using it much more heavily, reaching 12% of output tokens versus 5% for people managers.

OpenAI is becoming the test case for a workplace where humans supervise fleets of agents.

Inside OpenAI, Codex drove a huge rise in AI-generated work across departments, with median output tokens rising 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since Nov-25.

@rohanpaul_ai the job is becoming giving tasks and checking work

Codex adoption is still mostly technical in outside organizations, with Engineering at 27% and Data at 15%, while inside OpenAI it has become the default AI work tool across functions, reaching 99% in Engineering

Non-developer Codex adoption is exploding fastest outside OpenAI, rising 189x in organizations and 137x among individuals since Aug-25.

Codex started with developers but is now growing even faster among non-developers, especially in external organizations, where active non-developer users rose 189x versus 85x for organizational developers since Aug-25.

Codex is doing most of the generated work once adopted, reaching 99.8% of OpenAI output tokens, 63.3% in organizations, and 16.5% among individuals.

Here, "Organizational" means external companies using OpenAI business products, mainly Business and Enterprise accounts.

Codex is becoming a parallel work engine for heavy OpenAI users, where the top 1% run roughly 70+ agent-hours per day, while organizational and individual users still use it much more lightly.

Codex users are moving from one-off prompts to reusable agent workflows, with 26.6% of active users using at least one skill overall and 96% of OpenAI users doing so.

@heynavtoor and one day that also will be taken over 😃

@rohanpaul_ai this is the part people keep underweighting.

once agents do the first pass across functions, the human job moves upstream: define the outcome, set the constraints, review the edge cases, own the result.

that is a very different org design.

@rohanpaul_ai AI现在最大的问题是：说得比做的好听。真正能落地的项目，远比融资PPT里的少。 「OpenAI just released a paper showing how they are now seeing...」

@rohanpaul_ai Let’s brace ourselves for another layoff wave

@rohanpaul_ai 71 hours of agent work per day is wild. How do they handle error recovery when multiple agents are running concurrently? Debugging that async mess sounds like a nightmare.

@rohanpaul_ai The job is shifting from doing the work to assigning it and making sure it's done right.

the 99.8% number is loud, but the 5+ concurrent agents and 71h/day part is the actual pricing problem. once Codex becomes delegated labor instead of chat, monthly tiers feel less useful than token unit cost + upstream reliability. I broke down that route here

@rohanpaul_ai 99.8% of output tokens isn't 99.8% of the work though. The interesting question is who still writes the 0.2% that decides whether the rest ships or gets reverted. That sliver is the whole job.