OpenAI Reports Codex Produces 64% of Enterprise AI Output
Two new studies suggest agent use is spreading beyond engineering, while a widening usage gap separates OpenAI's most active business customers from typical adopters.
OpenAI published two enterprise-usage studies on August 12 that track a shift from conversational assistance toward delegated work. In the company's June data, Codex generated 64% of the combined output tokens produced by Codex and ChatGPT among enterprise customers, according to the official report.
The findings matter to business leaders evaluating agent deployments because they describe where usage is deepening, not merely whether employees have access. OpenAI reports that activity is spreading into legal, sales, recruiting and marketing, while the most intensive customer organizations are pulling further away from typical adopters.
The enterprise usage gap widened
OpenAI defines frontier firms as the top 10% of enterprise customers by monthly output tokens per active user. Typical firms sit between the 45th and 55th percentiles. By June, the frontier group generated 8.3 times as many output tokens per active user as the typical group, up from a 2.6-times gap in January.
That measure is a proxy for depth of use rather than proof of productivity. Longer agent tasks can generate far more output than short assistant conversations, so token volume captures workload intensity as well as adoption. It does not establish that the extra output improved revenue, quality or employee performance.
The same split appears in advanced features. OpenAI says 21% of weekly active users at frontier firms used Plugins and 19% used skills, compared with 9% and 3% at typical firms. These capabilities connect agents to repeatable instructions, company context and external tools, making them more relevant to workflows that end in completed work rather than an answer alone.
Codex use moved beyond engineering
Since February, OpenAI says weekly active enterprise Codex users increased 108-fold in legal, 41-fold in sales, 41-fold in recruiting and 26-fold in marketing. Engineering grew fivefold over the same interval. The different growth rates do not mean those non-technical functions now have more total users; they show how quickly adoption expanded from a smaller starting point.
The two studies draw on OpenAI customer activity, including a sample of more than 10 million messages. That gives the company a large view of behavior inside its own products, but it also limits the conclusions. The cohorts are defined by OpenAI, the figures are not independently audited, and companies using rival systems or internal models are outside the measurement.
What teams can take from the data
The practical signal is that access alone is not producing uniform adoption. Organizations reporting deeper use are also using more connectors, reusable skills and agent workflows. For buyers, the next useful checkpoint is not another seat count but evidence that these workflows complete valuable tasks under clear permissions, review and governance.
Status
Confirmed. OpenAI published the studies and the reported measurements. Internal confidence is medium because the usage figures, cohort definitions and interpretation come from OpenAI without independent corroboration.
Sources
Update note: Last reviewed 2026-08-12. We will revise this post if OpenAI publishes the underlying methodology or independent evidence tests the reported adoption patterns.
Sources
- OpenAI — enterprise AI adoption studies — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.