Nearly 98 percent of OpenAI’s employees now use Codex, the company’s AI coding agent, up from roughly 40 percent in August 2025, according to a paper the company published on Wednesday titled “The Shift to Agentic AI: Evidence from Codex.” The paper describes a fundamental change in how the company’s own workforce interacts with AI, moving from conversational chatbot use to autonomous agents that execute multi-step tasks. Every statistic in the paper, however, comes from OpenAI itself, a company with a direct financial incentive to promote the product it is measuring.

The headline numbers are striking. Active Codex users grew fivefold in the first half of 2026, and requests for tasks estimated to take eight or more hours increased nearly tenfold. OpenAI’s legal team generated 13 times more tokens in June than in November 2025, a figure the company presents as evidence that agents are penetrating departments far beyond engineering.

The growth among non-developers is where OpenAI’s narrative centres. Individual non-developer usage of Codex grew 137 times since August 2025, organisational non-developer usage grew 189 times, and internal non-developer adoption grew twelvefold. The company expanded Codex earlier this month with enterprise plugins connecting 62 business applications, and non-developers now make up roughly 20 percent of the platform’s five million weekly users, adopting three times faster than engineers.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

The paper frames this as evidence of a market-wide transition from chatbots to agents. OpenAI argues that the pattern visible inside its own company, where every department from legal to recruiting now treats Codex as a primary tool, previews how enterprise AI adoption will unfold broadly. The company points to external data showing Codex usage among organisations at roughly 17 percent and among individuals at under one percent, suggesting significant room for growth.

But the gap between internal and external adoption also raises questions about how representative OpenAI’s own workforce is. The paper does not address whether the company incentivises or encourages employees to use Codex, a relevant omission given that nearly universal adoption inside a company selling the product is not the same as organic demand. No independent third party has verified any of the usage figures.

The self-reporting problem extends to the productivity claims, where OpenAI says longer task requests and higher token generation prove that agents are handling more complex work. As The Register noted, faster code generation does not automatically translate into proportional productivity gains, because verification, testing, and deployment time may expand to absorb the speed improvement. The paper does not present data on whether the shift to agents has measurably improved output quality or reduced total time to completion.

The broader context is a race among AI companies to prove that agents, not chatbots, represent the next phase of the market. OpenAI merged ChatGPT and Codex under Greg Brockman in May, consolidating its product strategy around a single agentic platform ahead of a potential Q4 IPO. Anthropic’s Claude Code and Google’s Gemini are pursuing similar agentic strategies, making the competitive pressure to show adoption growth intense.

Meta’s internal experience offers a parallel data point. The company introduced a “Claudeonomics” leaderboard in April that tracks token consumption by team, turning AI usage into a visible performance metric. That approach, like OpenAI’s paper, measures input volume rather than output value, a distinction that matters when the companies reporting the numbers are also the ones selling the tools.

The paper is most useful as a signal of where OpenAI believes the market is heading, and where it wants investors to believe the market is heading before its IPO. The shift from chatbot queries to autonomous agent tasks is real and visible across the industry. Whether it is happening at the pace and scale that OpenAI’s self-reported data suggests is a question that only independent measurement will answer.