The marquee use case for AI agents was always supposed to be writing software. OpenAI's own usage data, leaked through its economic research arm, says otherwise.
There is a tidy story about AI agents that almost everyone tells, and OpenAI just quietly contradicted its own version of it.
The story goes like this: large language models are, at their core, coding tools. They write functions, fix bugs, and the marginal dollar of agent value flows to engineering teams shipping software faster. It is a clean narrative, and it is wrong about where the action actually is.
According to data surfaced by OpenAI's economic research team, internal Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025, per a report aggregated by Latent Space. Read that ordering again. Engineering, the department the entire product was named and built for, came third. Research, a knowledge-work function, grew more than twice as fast.
The same report notes that through August 2025, the average OpenAI worker spent less than 10% of their tokens on Codex. Then it inverted. Over six months, agent usage deepened and spread across departments that do not, in any traditional sense, write code for a living.
This is the most honest signal we have about where autonomous work is creating value, because it comes from inside the company with the most sophisticated agent users on earth, spending their own compute on their own problems. It is not a product launch. It is not a projection. It is revealed preference. And the preference is for agents that do the job, not agents that write the program that does the job.
