What happened
- On September 6, 2026, OpenAI announced that it had reached the goal it set last year of having an automated research intern by September 2026.
- The company defines that milestone as a system capable of carrying out well-defined research tasks under human direction, including work that would take an experienced person several days.
- According to its internal measurements, by mid-August the research organization was consuming 3.1 eight-hour agent workdays for every human workday. The threshold was only crossed after June 2026.
- The median researcher was spending more than $600 a day on inference at API prices; the top decile exceeded $7,000 a day.
Why it matters
- More than half of the successful four-to-eight-hour tasks included at least one human intervention. That percentage describes the real state of agents better than any adoption figure.
- For a small team in Chile evaluating whether to delegate processes to agents, the actionable number isn’t 3.1 but the spending: the $600-a-day median per person sets the floor for an operation that runs agents seriously.
- The ratio measures machine running time, not productivity. An agent can run in parallel, repeat itself, fail or require constant direction and still add hours to the numerator.
The number
3.1 agent workdays for every human workday, measured in eight-hour blocks.
Context
The same company set the goal, wrote the milestone’s definition, took the measurements and published the evaluation. There’s no external audit of the figures. OpenAI warns in its own text that the measurements are preliminary and don’t demonstrate progress in quality, model capability or scientific novelty. The documentation comes the same week the company placed Astra in its critical cybersecurity category.
What’s next
- The next stated goal is an automated AI researcher by March 2028.
- No independent evaluation has been announced to verify either milestone.
Bottom line
The indicators the company publishes (execution time, tokens spent, number of experiments) are the ones that are easy to measure. The ones that matter, such as whether the research advanced, still have no agreed way of being measured.
Sources
- Research acceleration: The view inside OpenAI — official OpenAI post, September 6, 2026
- OpenAI just hit a milestone on the road to self-improving AI — Help Net Security
Edited by Rodrigo Cornejo. How we select and verify the facts, in who writes.
