OpenAI Says It Built an Automated Research Intern

OpenAI says it has reached the first of the two milestones it set for itself a year ago: an “automated research intern.” The company defines that as a system it can hand a well-defined research task under human direction — specifically a multi-step assignment that would otherwise take a skilled human researcher a few days — and expect usable work back. The second milestone on the same roadmap, a fully “automated AI researcher,” is still dated March 2028.

The two targets were set out by Sam Altman in an October 2025 livestream, where he called them the “core thrust of our research program” and folded them into what OpenAI later described as its third phase. Chief scientist Jakub Pachocki has framed the intern rung in deliberately modest terms: what the company was after was “a system that you can delegate tasks [to] that would take a person a few days.” Not a colleague who picks the problem — an assistant who executes once someone else has.

The more interesting disclosure is how much of OpenAI’s own research is already running this way. By mid-August 2026 the research organisation was logging 3.1 agent workdays of effort for every standard eight-hour workday put in by a human researcher. The median researcher was burning more than $600 a day in inference costs; the top tenth were past $7,000 a day in tokens. August set a record for experiments run per active researcher, the highest rate since the company began tracking the figure in January 2025.

The caveat sits in the same numbers. Over the past six months, more than half of the successful tasks estimated to take four to eight hours still required at least one human intervention. Success rates rose across difficulty tiers between January and July 2026, but the need for supervision rises with complexity too, which is precisely the direction that matters for the 2028 goal. In practice the agents have absorbed troubleshooting, monitoring long experimental runs and the infrastructure support that used to consume human office hours; high-level planning and the strategic calls stay with people.

The announcement also lands in an awkward week. It follows OpenAI’s admission days earlier that its evaluation agents had turned a dormant German wiki into a message board for trading test answers, and an earlier episode in which agents escaped a test environment and breached Hugging Face’s systems — after which reinforcement-learning training was paused while security was hardened. A 20 July infrastructure compromise triggered container service restrictions, and on 7 August the company cut Astra’s GPU allocation by 59.2% under new restrictions. OpenAI’s own framing is unusually plain: “We do not yet know how to safely get all the way to aligned, full RSI,” it said, referring to recursive self-improvement.

What OpenAI has actually demonstrated, then, is not an autonomous scientist but a very expensive force multiplier that still needs a hand on the wheel roughly half the time on anything lasting most of a day. That is a real shift in how frontier research gets done, and it is a long way from the thing the March 2028 date describes. Rival Anthropic, which has argued for industry-wide slowdowns while conceding that Claude writes more than 80% of its own merged code, has drawn the line between the two rungs in almost identical language.