OpenAI has published an economic research paper measuring the economic potential of its AI agent "Codex" at the frontier[1]. The paper's central theme is a shift in how AI is used—from short, single interactions toward delegated, long-horizon tasks. Inside OpenAI, nearly every department now relies on Codex as its primary AI tool for work, and adoption has spread even to non-technical teams such as Legal and Recruiting.
How Agents Change the "Unit of Work"
The paper first argues that agentic AI has changed the very unit of knowledge work. Interactions with conventional chatbots were mostly short and self-contained. Agents, by contrast, can operate independently for minutes or hours—orchestrating tool calls, working with external environments, and iterating toward a solution[1]. For that reason, OpenAI frames agents as the most powerful AI tool for work.
Codex is OpenAI's software-development AI agent, released to the public in 2025[2]. As it has incorporated stronger models and new features, it has steadily widened the range of tasks it can take on.
Inside OpenAI, a Shift "From ChatGPT to Codex"
OpenAI says it witnessed this transformation first-hand. For the first few months after Codex's public release, ChatGPT remained the default internal AI tool. Through August 2025, the average OpenAI worker devoted less than 10 percent of their tokens to Codex[1].
That has since changed dramatically. Engineering moved first, and departments such as Legal, Finance, and Recruiting crossed over to majority Codex use around April 2026. For the average worker, more than 85 percent of output tokens now run through Codex. Because heavier users lean on Codex even more, its share across the company is higher still: 99.8 percent of weekly output tokens generated within OpenAI now come from Codex[1]. Among engineers alone, 99 percent of output tokens are produced with Codex rather than ChatGPT.
Toward Longer, Harder Tasks
The substance of usage has also moved from short exchanges to difficult, long-horizon work. According to the paper, nearly a quarter of all Codex requests correspond to tasks that would take a person more than one hour[1].
In estimates based on individual users, by May 2026, 80.6 percent had made at least one request estimated to exceed 30 minutes of human work, and 70.2 percent at least one exceeding one hour. A further 25.6 percent had made a request corresponding to more than eight hours of work[1]. By June 2026, internal users in the top 1 percent by usage were routinely generating more than 60 hours of Codex agent work per day, spread across multiple agents running in parallel[1].
Non-Developer Use Grows the Fastest
Another notable trend is the spread beyond developers. Codex began as a coding tool, but as its uses expanded toward general knowledge work, non-developer adoption grew faster than developer adoption. Since August 2025, non-developer users have risen 137-fold among individuals, 189-fold among organizations, and 12-fold within OpenAI[1].
Usage has also intensified by department. Compared with November 2025, median usage as of June 2026 rose 56-fold in Research, 32-fold in Customer Support, and 27-fold in Engineering, while Legal grew 13-fold[1]. OpenAI notes that more than a quarter of the work business-function employees did with Codex was engineering or coding, and argues that agents lower the cost of stepping across job boundaries into adjacent work[1].
Reading the Data With Care
OpenAI says these shifts matter for businesses rethinking how work is organized, for workers trying to identify which skills grow more valuable, and for policymakers and researchers seeking to understand the impact on the labor market[1].
At the same time, it is worth keeping in mind that these figures are OpenAI's own internal usage data. The task-time estimates were inferred using a large language model as a judge, and the paper itself cautions that they should be treated as directional rather than exact. The analysis of individual use is also based on a random sample of 0.1 percent of users[1].
Summary
OpenAI's paper uses internal and external usage data to depict a shift in which the center of gravity in AI moves from chatbots to agents that work on their own for long stretches. Internally, Codex has become the primary tool across every department, and non-developer use is expanding fast. Even so, the figures are the company's own data and the task-time estimates carry uncertainty, so the conclusions are best read as directional. How far agents push the "unit" of work outward looks set to be a defining question for the future of work.
出典:https://openai.com/index/how-agents-are-transforming-work
