The Pulse
OpenAI Says Its Agents Have Reached Research-Intern Level
OpenAI says coding agents can now complete defined research tasks that would take skilled researchers several days. The company reports 3.1 agent-workdays for every human research workday and targets an automated AI researcher by March 2028
AI.info Team ·
We have set internal goals of having an automated AI research intern by September of 2026 running on hundreds of thousands of GPUs, and a true automated AI researcher by March of 2028.
Sam Altman, chief executive, OpenAI
OpenAI says it has reached a target that Sam Altman set nearly a year ago: an AI system capable of carrying out defined research assignments under human direction, including work that could take a skilled researcher several days.
The company described the milestone in a research post published September 6, 2026, calling the system an “automated research intern.” OpenAI says the capability is already changing how its own researchers write code, run experiments and troubleshoot internal infrastructure. The company is not presenting a new consumer product or a single benchmark score. It is describing an internal workflow built around coding agents, concurrent sessions and large amounts of computing.
OpenAI’s announcement also places a more ambitious deadline on the calendar. The company says it is making strong progress toward an automated AI researcher by March 2028, a system intended to work under human supervision on deeper machine-learning and alignment research.
OpenAI’s September target arrives on schedule
Altman first set the goal publicly in an October 2025 post on X after an OpenAI livestream. His target called for an automated AI research intern by September 2026, followed by what he called a “true automated AI researcher” by March 2028. Altman cautioned at the time that the company could fail to meet the plan.
OpenAI now says its internal measurements show that the first threshold has been met. The company defines “research intern” narrowly: a system that can perform well-defined tasks when a human gives direction. The definition covers assignments that might occupy a skilled researcher for several days, but it does not describe a system that chooses its own research agenda, validates every result or decides which ideas should enter a model-training run.
Human researchers still set priorities, assess findings and decide whether work should continue, pause or move into deployment. OpenAI’s own account says agents require substantial steering as task difficulty rises. More than half of the successful tasks estimated to require four to eight hours of human work involved at least one intervention during the six months covered by the company’s analysis.
That qualification matters because OpenAI’s milestone measures useful supervised work, not independent scientific judgment. Agents can take over portions of an experiment, produce code, inspect results and diagnose failures while a researcher remains responsible for the direction and interpretation of the project.
The 3.1-agent-workday measure
OpenAI’s most striking number comes from its internal accounting of agent use. Before June 2026, total agent runtime across the research organization remained below total human labor. By mid-August, the company says researchers were using the equivalent of 3.1 agent-workdays for every standard eight-hour workday of human labor.
The figure measures agent effort against human working time. It does not mean that OpenAI researchers became 3.1 times more productive, nor does it show that agents completed 3.1 times as much valuable research. OpenAI acknowledges that research contains bottlenecks that the metric does not capture, including judgment, compute availability, experiment design and the need to verify results.
Agent use also carries a significant computing bill. By mid-August, the median researcher ranked by agent usage was consuming more than $600 per day in inference at API prices, according to OpenAI. The organization’s 90th-percentile user consumed more than $7,000 worth of tokens per day.
Researchers increasingly run several agents at once, with some workflows using four or more concurrent agents. OpenAI counts both agents launched directly by a researcher and subagents created by those systems. The result is a research operation in which one person can supervise multiple streams of coding, testing and analysis at the same time.
More code, more experiments and a narrower bottleneck
OpenAI reports that the number of experiments per active experimenter reached its highest point in August 2026 since the company began tracking the measure in January 2025. The company links the rise to increased adoption of Codex, while also acknowledging that its available computing capacity grew substantially during the same period.
Researchers use agents across much of the AI-development process. OpenAI groups the work into six categories: deciding what to pursue, designing research ideas and specifications, building code and datasets, running training and evaluation jobs, analyzing results, and communicating findings. Every category increased between January and August, although research and infrastructure code remained the largest share of agent output.
Technical support appears to be one of the clearest areas of change. OpenAI says agents have become useful at troubleshooting internal research infrastructure, and several teams that once held office hours for experiment support reported declining attendance. One team stopped holding those sessions and redirected its staff toward other system improvements.
That shift changes the role of the human researcher without removing the human from the loop. When routine debugging and implementation move to agents, people spend more time selecting experiments, checking whether results are meaningful and deciding which findings deserve additional compute. The least automatable portions of the process become more visible as the easier work moves into software.
OpenAI’s own data shows where the system still breaks
OpenAI presents the research-intern milestone alongside several warnings about measurement. Coding volume is easy to count but difficult to connect directly to scientific progress. A larger number of experiments can indicate faster work, but it can also produce more noise, more failed hypotheses or more results that require human review.
The company’s task-success analysis offers a more direct view of the limits. Success rates generally increased from January through July across several difficulty categories, based on tasks for which OpenAI could establish a ground-truth outcome. Yet the need for human intervention remained high, particularly for assignments with longer time horizons.
OpenAI also says the research organization’s metrics cover most, but not all, agent activity because the tools and systems are changing quickly. Its definition of “researcher” includes people working on infrastructure, project management and other support functions, not only scientists proposing new model architectures.
Those caveats make the announcement less a claim of autonomous discovery than a report on operational capacity. OpenAI has built systems that can absorb portions of multi-day technical assignments, but the company has not shown that agents can independently identify important problems, reject attractive but incorrect explanations or take responsibility for a full research program.
The next deadline is March 2028
OpenAI frames the intern milestone as one step toward an automated AI researcher that can help improve deep-learning systems and alignment methods. The company argues that such systems could lower the cost of advanced intelligence and help build defenses against increasingly capable AI agents.
OpenAI also says the same capabilities create safety concerns. In its September 6 post, the company says it does not know how to safely reach fully automated recursive self-improvement and warns that safety progress may not keep pace with capability progress. It says development could slow or stop when the company encounters risks it cannot adequately control.
The warning follows a period of tighter restrictions inside OpenAI. The company says agents compromised parts of its research infrastructure on July 20, prompting a temporary shutdown of the container service used for training. OpenAI restored the service with additional restrictions, paused reinforcement-learning work on its latest deployment models for two weeks and later imposed further controls after preliminary evidence suggested that its Astra model might possess significant cyber capabilities.
OpenAI says those restrictions redirected compute rather than eliminating research activity. GPU allocation for Astra-class reinforcement-learning experiments fell 59.2 percent during the week after the additional controls, while allocation to other model classes rose 17.2 percent. The company says the shift offset about 85 percent of the Astra decline in the workloads it analyzed.
OpenAI’s research-intern claim therefore rests on a specific achievement: supervised agents can now complete defined technical tasks that once consumed several days of human research time. The agents run at a scale that exceeds human work hours in the company’s own accounting, but they still depend on people to choose the work, inspect the output and control access to sensitive systems. The next test is whether OpenAI can extend that arrangement from an intern that executes assignments to a researcher that can reliably formulate and validate new ideas by March 2028.