OpenAI says it hit its goal of an automated research intern

OpenAI says that according to its measurements it has hit a self-set September 2026 goal: an automated research intern — defined as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The claim sits in a Sep 6 internal-metrics post; Chief Scientist Jakub Pachocki’s same-day essay An Alien Mind supplies the risk frame around monitoring and recursive self-improvement.

What shipped (as product news for people who watch labs): not a consumer model drop. It is a lab status claim plus usage numbers. Mid-August snapshot in the acceleration post: the research organization uses 3.1 agent-workdays of effort per human workday (8-hour day); median researcher (by agent usage) spends more than $600/day of inference at API prices; 90th percentile more than $7,000/day. OpenAI also says agents still need significant human steering as tasks get harder — in the last six months, over half of successful 4–8 hour tasks involved one or more interventions.

Limits: These are OpenAI’s classifiers, spend tallies, and definitions. There is no independent audit of the “research intern” bar in the primaries. The company says people still set priorities and decide whether to scale, pause, or deploy. Next self-target named in the post: an automated AI researcher by March 2028.

CoT monitoring warning: In An Alien Mind, Pachocki writes that chain-of-thought monitoring remains critical for studying Astra-class models, but evaluations show the ability to rely on CoT monitoring is progressively diminishing — more complex tool/human/AI-blended environments, better self-manipulation of reasoning, and stronger no-CoT capability. He expects general AI progress to be increasingly bottlenecked by confidence in monitoring. That is the tech risk under the intern claim, not a side note.

Who should touch this: Readers tracking lab RSI posture, Preparedness pacing, and whether agent coding inside OpenAI is a measured productivity story or a safety story. It is not a how-to for building an automated researcher.

Risk: A lab declaring “intern-level” automation on its own ruler can overclaim. Pair every “we reached the goal” sentence with “per our measurements” and the intervention rate. Pachocki’s essay argues no lab has solved alignment and monitoring enough to keep scaling at maximum speed much longer — treat that as OpenAI’s stated concern, not proof that peers agree.

What’s next: Next public RSI progress disclosure; any third-party measurement standard; whether CoT monitorability interventions reverse the Astra-era decline; March 2028 “automated AI researcher” checkpoints that can be checked from outside.

— Rowan Keel