Inside OpenAI's Codex Study: The Shift from Chat to Agents Is Happening Faster Than Anyone Expected

Important caveat: The paper explicitly warns that OpenAI is not representative of typical organizations. It calls itself a "frontier" environment where workers are highly familiar with models, usage costs are low, organizational buy-in is maximal, and informal knowledge sharing is common .

2. External Usage Growth: Rapid, but Uneven

External adoption of Codex is accelerating, but it looks very different depending on whether you are an organization or an individual user.

• Active Codex users grew more than fivefold in the first half of 2026, with the fastest growth coming outside the initial audience of software developers .
• Organizational users show significant uptake: Codex now accounts for 63.3% of output tokens among organizations active on both ChatGPT and Codex . The share of organizations using Codex grew from just above 0% in August 2025 to approximately 17% of all active organizational users .
• Individual (consumer) users tell a very different story. Only 16.5% of their output tokens come from Codex. Among consumers active on either platform in the last 28 days, fewer than 1% have used Codex at all .
• The fastest growth is in non-developer populations: non-developer individual users rose 137x and non-developer organizational users rose 189x since August 2025 .

3. Task Complexity Trends: Users Are Delegating Bigger Jobs

Perhaps the most striking finding is that users are not just using Codex for quick questions. They are handing over work that would take humans hours or even full days.

By May 2026, among sampled individual Codex users:

• 80.6% made at least one request estimated to exceed 30 minutes of human work.
• 70.2% made a request exceeding 1 hour.
• 25.6% made at least one request estimated to exceed 8 hours of experienced human work .

The share of individual users submitting a request estimated to require more than 8 hours grew nearly tenfold since the start of 2026 .

At OpenAI, power users are already operating at scale. Users at the 99th percentile now regularly generate more than 60 hours of Codex agent turns per day, distributed across multiple parallel agents . More than 10% of users manage three or more concurrent Codex agents each week, and 26.6% use "skills" — shared instruction sets for complex workflows . In total, the paper found that 24% of all Codex requests are for tasks estimated to take a person more than one hour .

4. The Broader Shift: From Conversation to Delegation

The paper frames the transition from conversational to agentic AI as a fundamental change in the unit of knowledge work.

"Agentic AI changes the unit of knowledge work from single interactions to delegated, long-horizon tasks," the paper states .

It distinguishes ChatGPT as primarily conversational (Q&A, advice) from Codex as agentic (delegated production: debugging, refactoring, drafting documents, analyzing data, coordinating communication). The paper notes this is not an absolute binary — ChatGPT has some agentic features, and some Codex use is conversational . Codex use is "strongly oriented toward delegated production" — users ask it to do work, not only to provide advice or information .

Originally built for software development, Codex now extends to drafting documents, data analysis, communication, and other knowledge-work tasks outside coding . Within OpenAI, over one-fourth of the work done with Codex by workers in business functions was engineering or coding — showing agents lower the cost of crossing task boundaries .

5. The Paper's Own Caveats and Limitations

The research is explicitly self-limiting about what its findings mean. These are not minor footnotes:

1. OpenAI is not representative. The paper states: "OpenAI usage is therefore not representative of the typical organization today. However, it provides a view of what agentic AI use may look like in the future, when adoption frictions are minimal" .
2. Conversational vs. agentic is a spectrum, not a binary. In the week before June 11, 2026, 60.3% of Codex turns and 21.9% of ChatGPT turns invoked at least one external tool — tool use itself is an imperfect proxy for agency .
3. Complexity estimates are model-generated. The task-duration thresholds (30 min, 1 hr, 8 hrs) are estimated by an AI model, not directly measured, and are based on a 0.1% random sample of individual users who opted to allow their queries to be used for training .
4. Adoption remains narrow overall. Among all consumers on consumer plans, fewer than 1% have used Codex — most AI users are still primarily chatting with bots .
5. External adoption is "lower and more uneven" than internal OpenAI adoption, and agentic tooling overall remains "much less broadly used than ChatGPT" .
6. The data is from one platform (Codex) and one company (OpenAI) — findings may not generalize to other agentic tools or organizations with different workflows, security requirements, and user familiarity levels .

Key Takeaways

• Agentic AI adoption is real and accelerating inside organizations, especially where cost and access barriers are low. But for individual consumers, chat-based AI remains the dominant paradigm.
• The biggest surprise may be the fastest growth happening outside software engineering — in legal, finance, recruiting, and other knowledge-work domains.
• Task complexity is increasing, with a significant minority of users already delegating full-day work to AI agents.
• The paper's limitations mean these findings should be read as a leading indicator, not a universal forecast. Most AI users are still chatting, not commanding agents.

OpenAI's research provides the most detailed look yet at how agentic AI is moving from a research concept to a practical work tool. The shift is happening — but where, how, and for whom depends heavily on context.