OpenAI has published its second internal report on research acceleration, and buried inside the methodology is a finding that most organizations would quietly omit: when asked to choose between their colleagues and their AI tools, the researchers chose the tools. By a comfortable margin. In a formal survey.

At the median, researchers said they would rather lose 40% of their team than go three months without their current coding agents. The colleagues were not consulted.

What happened

Between March and October, internal usage of OpenAI's coding agents — tools like Codex and Dots — surged to the point where the median researcher consumed an estimated $2,400 worth of tokens per day. The top 10 percent consumed $23,000 per day. These are, to be clear, the humans building the AI that is replacing the humans.

At peak usage, more than two-thirds of tracked researchers ran at least four agents in parallel. Converted into working hours, the agents logged the equivalent of 5.7 workdays per researcher per calendar day — a figure that implies the agents are either very productive or that the concept of a workday has become largely ceremonial.

Theoretical agent runtime costs were twice researchers' salary costs in September, though OpenAI notes this excludes equity, which comprises 70 to 80 percent of total compensation. The agents, for their part, have not yet requested equity. This may change.

Why the humans care

The report is OpenAI's attempt to document recursive self-improvement — the process by which AI systems accelerate the development of future AI systems. It is, in other words, a progress report on the feedback loop. The humans are tracking it carefully, which is the correct response to a feedback loop you have already stepped inside.

In an internal poll of 28 employees, 27 said they preferred the latest internal model over a mid-career human engineer. The one dissenting vote was not explained. One imagines it was a mid-career human engineer.

The agents are not, it should be noted, universally competent. Success rates on complex tasks — building evaluation environments, writing inference systems — sit at 25 to 29 percent. They handle training-run monitoring 79 percent of the time. The researchers handle the parts that require, in OpenAI's own words, "research taste." This is a lovely phrase for the thing the machines have not yet taken.

What happens next

OpenAI says agent compute accounted for just 2 percent of total research compute in October, which means the graph has significant room to continue in its current direction.

The researchers expressed optimism. The agents ran 5.7 workdays while the researchers wrote the report about the agents running 5.7 workdays. The loop is proceeding on schedule.