OpenAI Report: Internal Agent Daily Cost Reaches Twice a Researcher's Salary
nashnova research
An internal OpenAI report reveals that the daily API cost of agents driven by its researchers now runs at twice their non-equity compensation, with agents independently completing 46% of research tasks — AI's 'on-paper labor value' is structurally overtaking human pay.
How much does one researcher spend on agents each day?
Between September and October, the median researcher's daily agent spend rose from $1,500 to $2,400.
The top 10% of users went further: daily spend jumped from $11,000 to $23,000.
This means → a top researcher's one-day agent bill equals roughly half a month's salary for an average employee.
How many agents does one person run at the same time?
Using the benchmark of "8 cumulative agent-hours = one work-day," each researcher averaged 5.7 agent work-days per day — 11 times the prior baseline.
Over two-thirds of researchers run at least 4 agents simultaneously; more than 10% run at least 16.
In plain terms = one researcher sits at a desk while five or six AI "copies" work in parallel behind the scenes — some have over a dozen.
How much can agents actually finish on their own?
Agents now independently complete roughly 46% of research tasks that would take a human 2–4 hours; in March the figure was just 23% — it doubled in seven months.
For harder 4–8 hour tasks, the autonomous completion rate rose from 13% to 35%.
By task type: communication 69%, analysis 58%, but "building" only 43%; training-monitoring success hit 79%, while writing inference and serving code landed at just 25%.
This reflects a clear pattern: agents excel at executing defined procedures but remain weak where judgment and creativity are required.
How dependent are researchers on agents?
OpenAI surveyed 101 researchers: if you lost access to the best coding agents for three months, what would you trade to get them back?
The median answer: 40% of the team's headcount, or 30% of the team's compute.
This means → in researchers' own estimation, a good agent is worth more than nearly half their colleagues and a third of their computing resources.
Where do agents fall short?
In interviews, researchers said agents lack "research taste" — the ability to frame well-defined questions and judge the quality of code and data on their own.
In plain terms = agents can do the work and produce volume, but they don't know what work to do or whether it's any good — direction and quality judgment still require humans.
This also explains why agents handle 46% of tasks but the remaining 54% still need continuous human oversight.
Is safety investment keeping pace?
The report shows that among all research directions, safety and alignment saw the largest increase in headcount.
As AI accelerates its own R&D, human investment in controlling it is expanding in lockstep.
This reflects a core tension: the stronger and more widely used agents become, the greater the demand for "who watches the agents" — arguably the single most important variable to track in the current AI R&D cycle.
市场有风险,内容仅供研究参考,不构成投资建议。
