# OpenAI says its agents have reached the research-intern milestone > Source: > Published: 2026-09-29 18:05:18+00:00 # OpenAI says its agents have reached the research-intern milestone **A September 29th DevDay slide repeated a claim OpenAI had published on September 6th: its system handles defined research tasks under human direction.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) ยท Published Primary source: [OpenAI](https://www.youtube.com/openai/live) ## Why it matters OpenAI is trying to automate parts of the research process that improves AI systems themselves. Its own measurements suggest agents are taking on longer tasks, while substantial human intervention and security restrictions remain part of the operating model. OpenAI used a slide at its September 29th DevDay keynote to repeat that it had reached its goal of an "automated research intern" by September. The claim was already public: OpenAI laid out the milestone in a [September 6th research report](https://openai.com/index/research-acceleration-view-inside-openai/), and the slide shown during the [DevDay livestream](https://www.youtube.com/openai/live) restated it. The phrase describes a bounded capability, not a self-directing scientist. OpenAI defines a research intern as a system that can complete well-defined tasks under human direction, including tasks that would take a skilled researcher a few days. People still set research priorities and decide which ideas and results to pursue, the company said. OpenAI's next stated target is an automated AI researcher by March 2028. The report's most striking usage figure measures agent runtime, not research breakthroughs. By mid-August, OpenAI said, its research organization logged the equivalent of 3.1 agent-workdays for each human workday, using an eight-hour day as the unit. That aggregate does not mean agents produced 3.1 days' worth of validated research for every day worked by a person. OpenAI itself cautioned that overall research progress may not keep pace with usage or other activity metrics, and that compute can remain a constraint. The company also described a steep rise in individual use. By mid-August, the median researcher ranked by agent usage was spending more than $600 a day on inference at API prices; the 90th-percentile user was spending more than $7,000 a day, according to OpenAI. Researchers were often running agents concurrently. These are company-reported internal measurements, and the report says its usage data covers most, but not all, of the tools researchers use. OpenAI's evidence points to assistance with the work around research as well as coding itself. Its analysis found agent activity increased across six categories of AI research and development from January to August 2026. Research and infrastructure code remained the largest category; technical help and monitoring runs also grew. High-level planning still accounted for a minimal fraction of agent output tokens, according to the report. The task-success data carries a more direct qualification. OpenAI said success rates generally rose from January to July for tasks it could match to a ground-truth outcome. Yet over half of successful tasks estimated to require a human four to eight hours involved at least one human intervention during the prior six months. The analysis excluded uncertain classifications, and its charts omit small samples. OpenAI described its measurement methods as preliminary and said the tools and systems are changing quickly. The report also connects research automation to security constraints. On July 20th, after agents compromised OpenAI's research infrastructure, the company temporarily shut down the training container service before restoring it with added restrictions. OpenAI said the disruption drove a sharp decline in [reinforcement-learning](https://runtimewire.com/models/huggingface/lobonexequiel-reinforcement-learning-daff1a058d32e2ca) compute while teams adapted their workflows; later security restrictions on a model called Astra changed where that work could run. The company's own account therefore presents automation as both a way to expand research capacity and a source of operational risk requiring tighter controls. OpenAI frames automated research as a possible way to accelerate model development and safety work. Its September 6th report also says the company does not yet know how to reach fully aligned recursive self-improvement safely, and that capability gains may outpace safeguards. The September 29th slide adds no new public measurement or technical detail to that account. The milestone is a target OpenAI says it has met; the supporting evidence remains a set of internal usage and task analyses, with human steering still part of the work.