OpenAI reaches its automated research intern milestone
OpenAI says its internal measurements have reached the automated research intern milestone. Three publishers tie the mid-August ratio of 3.1 agent-workdays to that same announcement, but the figure measures agent runtime rather than completed research output. Human intervention in more than half of successful four-to-eight-hour tasks shows that the company's intern definition still preserves human direction.
Artificial Intelligence··Morning
What the 3.1 ratio measures
OpenAI announced on September 6 that, by its own measurements, it had reached the automated research intern goal set last autumn. Unite.AI and AI Weekly report that by mid-August the research organisation used 3.1 agent-workdays for every human workday. An agent-workday is based on eight standard hours of active runtime and includes agents launched directly by researchers as well as downstream subagents. The ratio therefore describes how much agent time the laboratory consumed; it does not show that research output rose 3.1-fold, that 3.1 separate tasks were completed, or that 3.1 equivalent results were produced.[1], [2]
The intern definition includes people
AI Understanding notes that the company defines its research intern as a system able to perform well-defined tasks under human direction. Those tasks can be long enough to take a skilled researcher several days, but people still decide what work to pursue, which results matter and whether scaling should continue. More than half of successful tasks in the four-to-eight-hour range involved at least one human intervention. The publication does not say whether an intervention was a brief clarification or a substantial rescue, so that percentage alone does not establish the system's degree of independence.[3], [1]
Usage cost and the next target
AI Weekly and Unite.AI report that daily inference use exceeded 600 dollars for the median researcher by mid-August and 7,000 dollars for the 90th-percentile user. These are not public prices for a research product; they are resource-use estimates that value internal consumption at API prices. OpenAI names its next goal as developing an automated AI researcher by March 2028. AI Understanding points out that the company has not published the task set, difficulty calibration or grading method. Without those details, the announcement demonstrates expanding internal use but does not allow an outside team to reproduce the measurement.[2], [1], [3]
Related columns
For more information on this topic, you can read the related columns.
