OpenAI says it has reached an internal milestone it set last fall: an automated research intern capable of carrying out well-defined research tasks under human direction that could take a skilled researcher a few days. In a September 6 post, the company said it is now working toward a more capable automated AI researcher by March 2028.
That is not a claim that OpenAI has built an independent scientist, nor an externally validated benchmark. The company’s definition is deliberately narrower. The reported system works on bounded tasks, people still choose the research priorities and make deployment decisions, and OpenAI says its measurements of agent-powered research remain preliminary. But the disclosure offers a concrete view of how frontier-lab work is changing as coding agents move from occasional assistance to a regular part of the research process.
What the automated research intern milestone means
OpenAI’s label describes a level of task completion rather than general scientific autonomy. The company says the system can tackle specific research assignments with human direction, including work that would otherwise occupy a skilled researcher for days. It does not say the system independently selects important questions, validates its own conclusions, or governs its own deployment.
That distinction matters because the phrase can sound broader than the evidence. Research is not one activity. It includes proposing ideas, writing and reviewing code, running experiments, interpreting results, detecting failures, deciding what to pursue, and communicating a finding. OpenAI’s post says agents are taking on more of the coding and operational work around that loop. It also says the least automatable tasks may become the limiting factor as other tasks are delegated.
The announcement therefore belongs less in the category of a conventional product launch than in a developing account of how a leading AI lab organizes its own work. It is a fresh development distinct from Unhyd’s recent report on GPT-6 Astra’s monitoring trade-off. The earlier piece examined a model’s deployment and supervision challenges. This disclosure is about the changing production system behind AI research itself.
OpenAI’s numbers show scale, not a finished answer
OpenAI reports that, by mid-August, the median researcher in its research organization was using more than $600 a day of inference at API prices. It says a high-usage researcher at the 90th percentile was using more than $7,000 a day. Before June, total agent runtime across the organization was below total human labor; by mid-August, the company says the organization was using 3.1 agent-workdays of effort for every human workday.
Those figures are striking, but they should be read with their limits intact. They are OpenAI’s internal measurements, not an independent audit of research productivity. Agent runtime is also not the same as human-equivalent output. A system can run in parallel, repeat work cheaply, or spend large amounts of compute on a task without producing the kind of insight that moves a field forward. OpenAI itself notes that its tools, methods, and research environment are changing quickly, which complicates direct comparisons.
The company does report evidence of a more practical shift. It says researchers are writing more code and running more experiments, while agent use is spreading to higher-level and longer-horizon tasks. Yet the post also says agents still need significant human steering as complexity rises: more than half of successful tasks estimated at four to eight hours involved at least one human intervention. That constraint is central to the story, not a caveat to skip past.
Human control is still the operating model
OpenAI says it does not yet know how to safely achieve aligned, full recursive self-improvement, or RSI. Its post argues that more capable automated research could help with safety and alignment, while also acknowledging that alignment and monitoring may not keep pace with capabilities. The company describes prior pauses and added restrictions in its research environment as examples of how safety controls can change the pace and mix of technical work.
For organizations watching the rise of workplace agents, the immediate lesson is more practical than speculative. Increasing automation changes the job of supervision; it does not remove it. The useful controls are still concrete: narrowly scoped permissions, clear ownership, action logs, meaningful approval points, and the ability to pause or reverse a workflow. Unhyd’s permission-first guide to AI agent security explores why those controls matter when an agent can act through workplace tools rather than merely generate text.
The next evidence worth watching will be more comparable measures of research outcomes, not just tool use. Are agents helping teams reach sound results faster? Which tasks transfer reliably? How much review is required? And can labs demonstrate that their safety work grows alongside the scope of the work they automate? OpenAI’s September update does not settle those questions. It makes them harder to avoid.
.jpg){kind=link}