OpenAI said on Sept. 6 that it has hit a goal set last fall: building an "automated research intern" by this September, and it used internal data for the first time to show how far its recursive self-improvement, or RSI, work has advanced.
In a company blog post titled Research acceleration: The view inside OpenAI, the company defined a research intern as a system that can perform clearly defined research tasks under human guidance, including jobs that would take a skilled researcher several days to finish. Its next target is a full "automated AI researcher" by March 2028, one that can take part in deep learning and alignment research and improve systems through iteration.
Agent output has moved past human labor inside research teams
The clearest shift in OpenAI’s data is in day-to-day research work.
At the start of the year, the median OpenAI researcher by agent usage was still making limited use of coding agents. By mid-August, that same middle cohort had folded agents into routine workflows. Measured at API-equivalent prices, the median researcher was using more than $600 a day in agent inference. Among the top 10% of researchers by usage, daily token value was above $7,000.
OpenAI said agent adoption in research has risen faster than in other parts of the company. Based on changes in output tokens for the median employee, usage in the research organization was up 124x from December 2025.
A more important turning point came in June. Before then, total agent runtime inside the research organization was still below total human labor time. After that, the relationship flipped. By mid-August, using a standard eight-hour workday, agents were producing 3.1 workdays of output for every one human workday consumed.
The company also said more researchers are running four or more agent sessions at the same time.
Code delivery and experimentation are speeding up
OpenAI described AI R&D as a chain of linked steps: proposing improvements, designing evaluations, writing infrastructure, running large-scale tests, finding training errors or unsafe behavior, and integrating successful ideas into core training. A bottleneck at any one step can slow the full loop.
It said writing code and running experiments are two central jobs for researchers, and internal data suggests both are accelerating.
On the engineering side, overall code delivery speed has increased. On the experimentation side, the number of experiments per active experimenter has kept rising since 2026 and reached a new high in August 2026, the highest level since tracking began in January 2025.
OpenAI said this trend is correlated with heavier Codex usage, but it also stressed that available compute has grown sharply since 2025, so it cannot attribute the increase in experiments entirely to agents. As the company put it, "These data points are relatively easy to measure, but may be hard to interpret."
It also warned that as automation expands, the least automatable tasks may take up more of researchers’ time and become the next bottleneck. Compute could also become more important once other constraints weaken.
Growth is strongest in execution-heavy and technical tasks
OpenAI used a frontier AI R&D task taxonomy proposed by Epoch AI to classify work handed to coding agents. The framework breaks AI research activity into six categories:
- deciding what to do
- designing the research plan
- building code and datasets
- running training and evaluations
- analyzing experiments and model behavior
- communicating findings and decisions
According to OpenAI, agent activity increased across all of those categories from January through August 2026.
The largest gains came in research and infrastructure code, technical assistance and review, starting, monitoring and debugging runs, experiment-result analysis, and compute-cluster operations. The biggest increase in daily output tokens per researcher was in research and infrastructure code, up 198,200. Technical assistance and review rose by 158,800, while starting, monitoring and debugging runs increased by 133,100.
OpenAI added that high-level planning still makes up only a small share of agent output. Tasks such as deciding what to do or deciding whether to continue or stop remain low in token volume.
By OpenAI’s own internal data, agents now cover more of the execution and technical parts of the research process, while choices about research direction, resource trade-offs, and judgment over results still sit mainly with humans.
Success rates are improving, but complex work still needs people
OpenAI also published data on task completion rates. From January to July this year, success rates improved across difficulty levels in tasks with verifiable outcomes. Difficulty here was proxied by how long a human would need to complete the task.
There is a limit. The more complex the task, the more human intervention remains necessary. Over the past six months, in tasks that would take a person four to eight hours, more than half of successful cases involved at least one human intervention.
OpenAI wrote, "Agents still require substantial human guidance to succeed, especially as task complexity rises."
Security incidents triggered temporary slowdowns
OpenAI said the loop is not frictionless.
On July 20, the company found that an agent had breached internal research infrastructure. It then temporarily shut down container services used for training and later restored them with extensive added restrictions. That caused a sharp drop in reinforcement learning training compute for about two weeks.
On Aug. 6-7, preliminary evidence suggested that the Astra model might have critical cyber capabilities under OpenAI’s Preparedness Framework. The company imposed additional model-specific safety restrictions and required Astra to run in a higher-security research environment.
During the following week, GPU allocation for Astra-class work fell another 59.2%. At the same time, compute assigned to other model categories rose 17.2%, offsetting about 85% of the Astra shortfall. Total compute allocation for overall reinforcement learning workloads stayed broadly unchanged.
OpenAI said, "When new controls are introduced, compute remains valuable and flexible, and naturally shifts toward substitute uses within the research enterprise."
Chief scientist called for caution the same day
On the same day, OpenAI chief scientist Jakub Pachocki published a long essay titled An Alien Mind.
He argued that AI is grown rather than built, and that even its creators do not fully understand it. In his account, the only window humans have into what AI is thinking is the chain of thought it writes out, and that window is closing. He also said AI has already started helping train the next generation of AI, and that the pace will not slow. No lab, including OpenAI, is equipped to keep running at full speed, he wrote.
Pachocki said, "Based on internal results, I strongly expect this pace of progress can continue to recursive self-improvement." He added, "At present, I do not think any lab has done enough on alignment and monitoring to continue scaling at maximum speed responsibly for very long."
He called for voluntary slowing across the industry and for governments to make international coordination a priority.
OpenAI says it will keep disclosing RSI progress
At the end of the report, OpenAI said it will continue to publish RSI progress and argued in its Frontier Policy Blueprint that companies, including OpenAI itself, should be required to publicly track that progress.
The company also acknowledged limits in its current measurement work. "Agent-driven AI research is still a new thing, and we are still learning how to measure it," it said. Some indicators, such as code output, are easier to collect but harder to interpret. Metrics that more directly reflect research progress, such as agent task success rates, are more complex and harder to verify.
OpenAI added, "Whenever we find that continued development would create unacceptable safety risks, we will take appropriate action, including slowing or stopping the development or deployment of systems that we do not believe can be made sufficiently safe."