OpenAI says its agents now outwork its humans 3-to-1

OpenAI's research organization logged 3.1 agent-workdays of automated effort for every one workday of human labor, according to the company's own operational snapshot published this month (OpenAI). The figure is self-reported, unaudited, and rests on an "agent-workday" unit the company has not publicly defined.

The same update describes agents taking on infrastructure work that previously needed specialist support, and frames the lab as still human-led: people set research priorities and judge results, while agents give the team more capacity to pursue promising ideas (ANI via Big News Network).

OpenAI bundled the metric with a claimed internal-model contribution toward a Navier-Stokes solution, a 20% drop in serving cost on GPT-5.6 "Sol", and a custom inference accelerator called Jalapeno. Each carries its own verification load: the math claim's status, the cost figure's benchmark conditions, and whether the chip is in production traffic or benchmark-only.

What 3.1 actually counts, which workflows shifted, and whether headcount fell as agent work rose remain unanswered in the public post.