Yesterday’s research-acceleration report says OpenAI was using 3.1 agent-workdays per human workday by mid-August. The median researcher consumed inference worth over $600 a day at API prices—a valuation of usage, not disclosure of OpenAI’s actual cost.
The more revealing qualification: over half of successful tasks in the four-to-eight-hour category required at least one human intervention. The task analysis excludes uncertain outcomes. Growing experiment counts also coincide with more available compute, so the study does not isolate an agent-driven productivity multiplier. OpenAI, September 6
For research and risk technology, I would treat this as an argument for changing the measurement system. Track accepted experiments, reproducible results and reviewer time—not just generated code or concurrent sessions. Faster implementation can move the queue to specification, evaluation and approval without shortening the complete research cycle.
These are useful firsthand operating measurements, but still self-reported vendor evidence. “Three agent-days” is a resource-consumption statistic until someone demonstrates what survived review.