OpenAI Research Intern Milestone

OpenAI says it has reached a goal it set for itself last fall: an automated "research intern" that can take a bounded research task, work on it under human direction for days, and hand back something useful.

The number behind it is the interesting part. As of mid-August, the org is running at 3.1 agent-workdays of effort for every human workday, using an 8-hour day as the baseline. That is not a demo statistic. That is throughput.

This is not the AGI headline. It is a capacity headline. Research-heavy teams have always been bottlenecked by how many skilled hours they can put against a backlog of open questions. If that constraint loosens even partially, the bottleneck moves somewhere else: review quality, judgment on which questions matter, and the discipline to keep a supervised system pointed at the right objective.

OpenAI's own framing keeps this bounded on purpose. This is a system that takes a defined objective and returns work for human evaluation, not one setting its own research agenda. The next checkpoint they are aiming for is a genuinely autonomous AI researcher by March 2028.

For any organization that competes on R&D velocity, the question is no longer whether this kind of leverage arrives. It is whether your review and evaluation process can absorb three extra days of work per person without becoming the new bottleneck.

If your team suddenly had three days of research thrown back at you for every one day your people put in, would your process be ready to use it well?