OpenAI announced on September 7, 2026, that it had achieved the goal it set of reaching the level of an “AI research intern” by September of the same year. According to the company’s definition, this description refers to a system capable of carrying out well-defined research tasks under human direction and completing work that could take a qualified researcher several days.
Reaching this level does not mean that the system has become an independent researcher. Humans remain responsible for setting research priorities, selecting ideas worth pursuing, and deciding whether to scale up or shut down systems. OpenAI had presented this path in two stages in October 2025, while its next goal is to reach the level of an “AI research scientist” by March 2028.
Coding Agents Move to the Center of Research Work
Data shared by OpenAI reveals a rapid shift in how its research teams work. At the beginning of 2026, researchers’ use of coding agents was relatively limited, but by mid-August the average researcher was using these systems regularly as part of their daily workflow.
The company estimates that the daily inference cost of the agents used by the average researcher exceeds $600 when calculated according to API prices. The agents’ total operating time also rose from a level below human working time before June 2026 to approximately 3.1 agent workdays for every human workday within the research organization in mid-August.
Researchers run multiple agents simultaneously, while these systems have begun performing tasks including addressing research infrastructure problems, writing code, analyzing experiments, and providing technical support. OpenAI says this use has helped researchers produce code more quickly and conduct more experiments. In August 2026, the average number of experiments per active researcher reached its highest level since measurement began in January 2025, with the company linking this to increased use of Codex and greater available computing capacity.
What Has Changed in Practice?
Agents are no longer limited to writing research and infrastructure code; their role has expanded to include technical assistance, experiment monitoring, and tasks that extend over longer periods. But this expansion does not amount to full autonomy. Over the past six months, more than half of the tasks that lasted between four and eight hours required at least one instance of human intervention.
Accordingly, the “research intern” described by OpenAI is an advanced assistant operating within a scope defined by a researcher, not a system that creates an independent research agenda or decides on its own which ideas should be developed.
The Next Step and Questions of Control
OpenAI links its more distant goal to the idea of increasing automation of AI research, ultimately enabling systems to help develop new systems and accelerate safety and alignment research. The company points to potential uses such as developing defenses against dangerous AI agents and protecting critical infrastructure.
However, the company does not consider rapidly achieving broad recursive self-improvement, or RSI, an inevitable goal. It says that continued progress in this direction will depend on humans’ ability to retain control and on safety procedures advancing in parallel with model capabilities. This remains the most important limitation in the announcement: the figures presented describe OpenAI’s internal experience and do not, by themselves, prove that automation will scale at the same rate outside its environment or without human intervention.