A company measure, not a headcount

OpenAI has published an unusually detailed account of how its own researchers use coding agents. Its headline number is easy to misread. By mid-August, the company says its research organisation was using 3.1 agent-workdays for every human workday. An agent-workday is not an employee or a person. It is agent runtime converted into an eight-hour unit.

The figure still says something real about the working environment inside a frontier lab. OpenAI says that, before June, its total agent runtime was below total human labour in that organisation. By mid-August, it says that relation had reversed. Researchers were running more concurrent sessions, writing more code and carrying out more experiments.

But runtime is an input. It is not the same as a research result, a reliable experiment or a safe model. OpenAI itself says AI research has many bottlenecks and that the pace of progress is unlikely to mirror any one of these internal activity measures.

The target is a research intern, with a person still directing the work

OpenAI says it has reached a target it set last autumn for an automated research intern. The company defines that as a system that can perform well-defined research tasks under human direction, including tasks a skilled researcher could take a few days to complete. It says it is working toward an automated AI researcher by March 2028.

That definition is narrower than a researcher who independently chooses a question, judges surprising results and changes a field's direction. OpenAI says people still set research priorities, decide which ideas to pursue and choose whether to scale, pause or deploy a system.

The company also reports a limit that matters. In its analysis of tasks with a known outcome, more than half of successful tasks estimated to take four to eight hours involved at least one human intervention. The systems are doing more work, but human steering has not disappeared when the work becomes harder.

More code and experiments are not the same as faster science

OpenAI reports that the number of experiments per active experimenter reached a high in August, alongside increased use of coding agents. It also says its available compute has grown, which is an important qualification. More experiments can come from better tools, more hardware, better infrastructure, different research priorities or some mixture of all four.

The post makes a useful distinction between tasks that are easy to count and progress that is harder to judge. Code produced, experiments launched and agent tokens consumed can all be measured. Whether those activities lead to a durable insight, a robust result or a safer system takes longer to establish.

That is why the number is best read as a work-practice signal. It shows that one leading lab is reorganising around agentic tools. It does not show that every field, organisation or research problem is about to become three times faster.

The acceleration is happening alongside new limits

OpenAI's update comes after its August decision to slow some frontier-model work while it changed security and monitoring practices. The company said it paused two weeks of reinforcement-learning training for its latest intended-for-deployment models, then brought some workloads back under tighter controls while others stayed paused.

OpenAI says its expanded monitoring can add roughly 20 percent of the inference compute being monitored, depending on the workload. That is a reminder that stronger safety work can itself consume capacity and slow a simple story about automation racing ahead without friction.

The company argues that automated research could also improve safety and alignment work. That is a plausible goal, not evidence that the hard safety problems are solved. OpenAI says it cannot assume that progress in alignment and safety will keep pace with general capabilities.

What is confirmed, what OpenAI says, and what is open

Confirmed: OpenAI published its research-acceleration account on 6 September 2026. It reports the 3.1 agent-workday figure, defines its automated research intern target and describes human interventions on some longer tasks. Its August security post separately records a training pause and new monitoring controls.

OpenAI's claims: agentic tools are meaningfully accelerating research work inside the company; the organisation has reached its internal research-intern target; and future automated research could help with alignment and defence. Those claims rely on OpenAI's own measurements and interpretation.

Open questions: how its measures translate to research quality, whether another lab would see similar results, how much of the observed change comes from added compute, what independent scrutiny of the methods will show, and whether safety and human control can improve at least as quickly as the agents' capabilities.

Sources

  1. OpenAI — Research acceleration: The view inside OpenAIPrimary OpenAI publication, 6 September 2026. Source for the internal agent-runtime metrics, research-intern definition, task-intervention analysis, methodology and stated limitations.
  2. OpenAI — Pacing model development in an era of cyber-critical capabilitiesPrimary OpenAI publication, 18 August 2026. Source for the reinforcement-learning pause, research-environment controls, monitoring design and stated monitoring-compute overhead.