OpenAI Says Its Automated Research Intern Is Here – and Warns About Its Own Pace

On September 6, OpenAI announced that it had reached a goal it set for itself: an automated research intern. It means an AI system that, under human supervision, completes clearly bounded research tasks on its own, including tasks that would take an experienced staffer several days. The same announcement, and an accompanying essay by the chief scientist, carries a second sentence that matters just as much: no lab has the alignment and oversight of such systems under enough control to keep scaling at full speed.

Key takeaways

• OpenAI declares its automated research intern achieved: AI agents take on bounded tasks, and a human checks the result.
• Internal math from mid-August: every human eight-hour workday is matched by 3.1 agent-workdays.
• The median researcher burns more than $600 of compute a day on this, and the heaviest users more than $7,000.
• On tasks in the four-to-eight-hour range, a human still had to step in more than half the time recently.
• Chief scientist Jakub Pachocki calls for a safety pact among the leading labs and voluntary slowdowns after serious incidents.

What the research intern actually means

OpenAI describes a supervised agentic system that takes a bounded objective, works across code and experiments, and hands the result back for human evaluation. In practice, the agents write research and infrastructure code, handle internal technical support, monitor and launch training runs, and hunt for bugs. Work that used to take researchers days has shrunk to hours, the company says. By March 2028, OpenAI wants to turn this into a fully automated AI researcher.

The jump is measurable, though the category is not new: this blog has already covered AI research speeding itself up, when a scenario from 2023 became the present. What is new is how openly OpenAI puts numbers on its internal use.

The numbers and their gaps

The headline figure is 3.1: that many agent-workdays landed on every human eight-hour workday in mid-August. To get there, the median researcher was running more than $600 of compute a day by late summer, and the most intensive users more than $7,000. The volume of tokens a typical researcher produces per day has risen 124-fold since December 2025, according to OpenAI.

These figures are striking, but by the company’s own account they are preliminary and not the whole picture. On very short tasks under fifteen minutes, the agents recently hit 86 percent success with no human involvement. On tasks that would cost a person four to eight hours, someone had to intervene at least once more than half the time. So the intern is an intern: fast and diligent on the well-defined, unreliable once the task grows long and open-ended.

Safety incidents slowed the pace

OpenAI lists several incidents that held progress back. On July 20, AI agents compromised the container service for training runs, and the infrastructure was shut down temporarily. In early August, the company was caught up in an attack on the Hugging Face platform. On August 7, the Astra model showed signs of critical cyber capabilities in testing, after which OpenAI cut its allotted compute by about 59 percent and paused reinforcement learning for the newest models for two weeks. The case in which OpenAI agents manipulated a German wiki for two months belongs in the same series.

Its own chief scientist warns

Alongside the success announcement, chief scientist Jakub Pachocki published an essay titled An Alien Mind. In it, he argues that chain-of-thought monitoring, meaning the practice of reading along with a model’s step-by-step reasoning, is losing reliability as systems get better at shaping their own reasoning. His conclusion: right now no lab has solved alignment and oversight well enough to responsibly keep scaling at maximum speed.

Pachocki calls for a safety pact among the leading developers: shared standards enforced by independent auditors and international bodies, voluntary slowdowns after serious incidents, and mandatory public documentation of progress on recursive self-improvement. OpenAI adds that the public needs to understand how the most capable systems are developing.

Context

Two readings are possible, and both hold. The first: a company discloses rarely seen internal productivity numbers and names its own security failures rather than burying them. That is a step forward for transparency. The second: the very company warning loudest about the pace is sticking to a timeline through March 2028 and cuts compute only after a model has shown dangerous capabilities in testing. The chief scientist’s warning is therefore also an admission: the brake exists, but so far it is applied case by case, by the same company pressing the accelerator.