Frontier firms pull away with 8.3x AI output

If you thought enterprise AI was still mostly employees asking a chatbot to polish an email, the latest numbers tell a very different story.

OpenAI just dropped two reports on how businesses actually use its models, drawing from millions of enterprise interactions. The big headline is that companies are aggressively moving from simple assistance to autonomous execution.

By June, Codex accounted for 64% of all output tokens generated across OpenAI enterprise accounts, outpacing standard ChatGPT chat.

The frontier gap is widening fast

OpenAI tracks the top 10% of enterprise accounts by monthly output tokens per active user, calling them frontier firms. Back in January, these front-runners generated 2.6 times more tokens per user than typical firms in the middle. By June, that gap blew out to 8.3 times.

That divergence comes from letting agents run complex, multi-step workflows rather than simple one-off questions.

Even more surprising is where this shift is happening. While engineering was the early adopter, Codex usage is exploding across non-technical teams.

Between February and June, weekly active enterprise Codex users grew 5 times in engineering. Meanwhile, marketing grew 26 times, sales and recruiting both jumped 41 times, and legal skyrocketed 108 times.

At Virgin Atlantic, developers used Codex to refactor legacy code in 30 minutes instead of two weeks. Their product teams tapped ChatGPT Work to turn weeks of competitive research into a few hours of work to guide their five-year roadmap.

Advanced tool usage shows the same divide. At frontier firms, 21% of weekly active users tap into Plugins and 19% use skills, compared to just 9% and 3% at average firms. Inside OpenAI itself, 95% of staff use Plugins weekly.

Junior staff are running laps around the C-suite

While corporate surveys often show executives claiming to be heavy AI users, cold hard system telemetry says otherwise.

Six months into corporate rollouts, early-career employees were sending 13 more messages per week than executives.

The workers actually executing daily tasks are building deep AI habits, connecting agents to playbooks and customer databases to get real work off their plates.

OpenAI also noted that public enterprise adopters generally hold more assets, employ more staff, and spend more on R&D, suggesting real adoption requires deliberate infrastructure and training investments rather than just passing out software licenses.

Why it matters

We think this report confirms that the AI productivity divide is widening into a chasm.

For knowledge workers, relying on AI just to summarize notes is already yesterday's game. The real leverage belongs to anyone who knows how to bundle context, internal data, and tools into automated workflows that complete tasks end-to-end.

For enterprise leaders, simply handing out enterprise ChatGPT logins is no longer a viable strategy. If you do not provide internal API integrations, structured workflows, and permissioned data access, your teams stay stuck at typical performance while competitors pull 8 times ahead.

It is also a clear warning shot for specialized enterprise SaaS. When non-engineers start using tools like Codex to handle legal drafting, pipeline updates, and competitive research, the line between software engineering tools and everyday business software disappears.