OpenAI researchers now run 3.1 agent-workdays per human workday

OpenAI's latest post is an unusually detailed look inside its own research operation, offering concrete metrics on how coding agents have changed daily work. The company frames this as part of a broader push for transparency, arguing that the public needs to understand how the most capable systems are developing to have a meaningful voice in how AI progresses. The data shows a clear shift: median researcher agent usage has gone from modest to daily, and total agent runtime has surpassed total human labor in the research organization.

A key subplot is how safety restrictions are shaping research. Following the Hugging Face incident and evidence about Astra's potential cyber capabilities, OpenAI imposed new controls that caused a sharp decline in Astra-class GPU allocation. However, the data shows compute was redirected to other model classes, offsetting about 85 percent of the decline. This suggests that when new controls are introduced, compute remains flexible and gets channeled into alternative uses, a pattern OpenAI itself flags as important for ongoing conversations about training and safety.

The stakes are significant for OpenAI's stated mission of building aligned artificial general intelligence. The company maintains that people still set research priorities and make deployment decisions, but the accelerating pace of agent-driven research could make this harder to sustain. OpenAI says it does not yet know how to safely get all the way to aligned, full recursive self-improvement, and it acknowledges that more capable systems can become harder to monitor. The coming months will test whether its raised safety and alignment standards can keep pace with the velocity its own agents are enabling.