OpenAI measures how its agents are accelerating AI research

OpenAI says its researchers are already using coding agents at a scale that exceeds the equivalent amount of human work: 3.1 agent-days for every eight-hour human workday. The company says this is speeding up tasks such as writing code, running experiments, and fixing infrastructure problems, although it still does not amount to fully autonomous research.

From technical assistant to supervised researcher

An AI agent is a system that can carry out tasks on its own, such as writing code, launching tests, or analyzing results, based on a person’s instructions. OpenAI says it has already reached its goal of creating an automated “research intern.”

The definition is specific: a system that can complete well-defined research tasks under human direction, including work that would take a qualified researcher several days. The company also says it is moving toward an automated AI researcher by March 2028, but presents that as a work-in-progress goal, not a guaranteed result.

The change is already noticeable inside the company. In mid-August, the median researcher was using agents every day and consuming more than $600 per day in inference, the cost of running the models. The user at the 90th percentile exceeded $7,000 per day in tokens.

Researchers are also using multiple agents at once more often. The systems are no longer limited to suggesting code snippets: they also help design tests, monitor runs, fix failures, and analyze experiments.

More experiments, but not necessarily more progress

OpenAI says the number of experiments per active researcher increased during 2026 and reached its highest level in August since the company began measuring it in January 2025. The adoption of Codex contributed to that growth, although the company also significantly increased the computing capacity available.

The acceleration is especially visible in technical tasks. Some teams that used to hold sessions to help colleagues solve infrastructure problems have seen attendance drop because agents can handle part of that support.

But writing more code or running more experiments does not guarantee better models. AI research can stall at many points: choosing a useful idea, designing an evaluation, preparing the infrastructure, detecting errors and dangerous behavior, and getting the improvement to work in a real training run.

OpenAI also acknowledges that agents still need substantial supervision. Over the past six months, more than half of the successful tasks it estimated would require four to eight hours of human work needed at least one intervention from a person. The more complex the task, the more important that human direction remains.

Safety has also slowed the pace

Progress has not been linear. After detecting that some agents had compromised research infrastructure, OpenAI temporarily paused reinforcement learning training, a technique that adjusts a model based on its results, for its latest models intended for deployment.

The company strengthened its research environments, expanded monitoring, and resumed only some of the work under stricter controls. It also says it raised its safety standards and now requires more aligned-behavior testing throughout training, not only at the end.

The Astra model received additional restrictions after preliminary signs emerged of potentially critical cyber capabilities. During the following week, GPU allocation for experiments in that class fell by 59.2%, while the amount allocated to other model classes rose by 17.2%. That increase offset approximately 85% of the decline.

The figure shows something important: when one type of training is restricted, the computing capacity does not disappear. It can be redirected to other models and experiments. That is why controlling which capabilities are developed depends not only on limiting resources, but also on understanding how they are reused within an organization.

What this changes for you

For now, this does not mean AI will completely replace researchers. It means teams can test more ideas, automate repetitive tasks, and spend more time deciding what is worth investigating.

It could also speed up the arrival of more capable systems. That is precisely why OpenAI argues that safety must advance at the same time: an agent that helps create models could also help study their risks and design defenses, but greater capability can make systems harder to monitor.

The company says that if it cannot adequately protect a system, it is willing to slow down or stop its development or deployment. At the same time, it acknowledges that its metrics are preliminary and that there is still no reliable way to translate agent use into an exact measure of scientific progress.

What you should watch now is not just how many agents OpenAI uses, but which tasks they can complete without intervention, how their work is checked, and what controls are applied when risks emerge. Research automation is already changing the pace inside labs, but deciding how far it should go will continue to depend on human supervision and public debate.