What does 3.1 agent-workdays per human workday count?

What the ratio counts

OpenAI reports that as of mid-August 2026 its research agents supplied 3.1 agent-workdays of effort per human workday, and treats that as reaching the automated research intern goal announced last autumn. The denominator is a human workday; the numerator is agent effort consumed in the same period. A ratio built that way rises when researchers run more agents, and it rises whether or not the extra agent time produced anything a researcher kept.^[1]

The spending figures in the same post have the same shape. A median researcher was past 600 dollars a day on inference by mid-August and the 90th percentile user past 7,000 dollars a day. Those are measures of resource use, and they belong to the numerator too. Nothing in them says how much of the work came back finished, which is the quantity the phrase automated research intern is meant to describe.^[1]

The counter-figure in the same post

Over 50 percent of successful 4-8 hour tasks required human intervention. The milestone is defined as research tasks carried out under human direction, so the intervention figure sits inside the definition rather than beside it: the same post that announces the intern also reports that on the longest tasks it counts as successes, a person stepped in more often than not. Where direction ends and completion begins is the whole question, and the post does not draw that line.^[1]

There is a reading in which the figure is harsher than the reality. If any researcher message during a run is logged as intervention, then a single clarification counts the same as a rescue, and the true share of unaided work is higher than 50 percent suggests. The reported rise in success rates across difficulty buckets between January and July 2026 is consistent with that reading. It is also consistent with the opposite one, because the task set, its difficulty calibration and the grading rule are not published, so the trend has no outside baseline to sit against.^[1]

The measurement that would settle it

A single number would do most of the work here: the share of tasks the agents completed with no researcher message at all, reported against a task set someone outside the laboratory can inspect. Add the grading rule and the difficulty calibration and the claim moves from an internal measurement to a result. Without them, 3.1 remains an honest description of how much agent time a research organisation now spends, which is a real change, and not a description of what that time returned.^[1]

OpenAI publishing the task set, the grading rule and an unaided completion rate would turn the milestone into something testable from outside. The company has given itself a schedule already, describing strong progress toward an automated AI researcher by March of 2028, and a schedule is easier to state than a measurement. The thing to watch by the end of this year is whether a document naming the tasks and the grading rule appears, rather than another ratio.^[1]