AI 代理與研究自動化
OpenAI Says Research Agent Work Hours Have Surpassed Human Hours Threefold, but Output Metrics Still Do Not Equate to Scientific Progress
OpenAI has released internal data showing that, as of mid-August, every human researcher workday corresponded to 3.1 agent workdays, reaching the company’s internal “automated research intern” milestone. The data suggests that agents are increasing experimental and engineering throughput, but most long-running tasks still require human intervention, while growing compute capacity makes causal effects difficult to disentangle.
OpenAI has provided its first relatively comprehensive account of how coding agents are changing its research workflows. The company defines an “automated research intern” as an agent that, under human guidance, can complete a clearly scoped task that might otherwise take a skilled researcher several days—not one that independently chooses research directions or decides whether to scale up training, pause experiments, or deploy models.
As of mid-August 2026, the aggregate runtime of agents across OpenAI’s research organization was equivalent to 3.1 agent workdays for every human workday. Based on API pricing estimates, the median researcher by agent usage consumed more than $600 in inference resources per day, while usage at the 90th percentile exceeded $7,000. Researchers are also increasingly launching four or more agents simultaneously to parallelize debugging, experimental infrastructure, monitoring, and analysis. In August, the number of experiments per active experimenter reached its highest level since tracking began in 2025, although high-level planning still accounted for only a small share of agent output.
These figures should not be interpreted directly as a 3.1-fold increase in scientific productivity. An “agent workday” measures execution time, not peer-validated discoveries, and the growth in experiment volume coincided with increased compute availability. OpenAI estimated success rates only for tasks with available ground truth, and over the past six months, more than half of the successfully completed four- to eight-hour tasks required at least one human intervention. The relevant classifiers, internal task distribution, and raw data have not been made public.
For research platform teams, the more meaningful trend to track is the shift in bottlenecks: as coding and debugging become heavily parallelized, GPU quotas, experiment design, result interpretation, and safety reviews may become the new critical path. External organizations seeking to reproduce these gains would need to measure valid experiments, the number of human corrections, failure costs, and final research outcomes—not merely tokens, runtime, or commit volume.