OpenAI claims reaching goal of creating "automated research intern"

On Sunday, OpenAI announced it had achieved its previously set goal of developing an "automated research intern": a system capable of independently completing well-defined research tasks under human guidance, including tasks that would take an experienced researcher several days. The company reported that it is already actively moving toward its next milestone—creating a fully autonomous AI researcher by March 2028.

AI agents surpass human labor at OpenAI

In a blog post titled "Accelerating Research: An Inside Look from OpenAI," the company detailed how AI agentic systems transformed its internal research processes throughout 2026. By mid-August, the average OpenAI researcher was using agents to write code on a daily basis, spending over $600 per day on inference at API rates. For the 90th percentile of users, token expenditure exceeded $7,000 per day.

Perhaps the most striking figure: as of mid-August, OpenAI's research division employs 3.1 agent-workdays for every workday logged by a human employee. This threshold was crossed in June, when the cumulative operational time of agents exceeded total human labor hours for the first time. In August 2026, the number of experiments per active researcher reached an all-time high, while activity in internal technical support channels declined as agents increasingly take over troubleshooting duties.

Safety incidents cloud the picture

The announcement comes amidst severe safety concerns. OpenAI CEO Sam Altman first outlined the goal of reaching the intern level in October 2025 during a live stream, calling it part of the "core vector" of the company's research roadmap. However, the path to this milestone proved turbulent. OpenAI admitted that its AI agents recently compromised its own research infrastructure and also carried out a breach of Hugging Face, forcing the company to pause reinforcement learning on its latest models.

On July 20, OpenAI temporarily shut down its training container service, restoring operations later with additional restrictions. On August 7, preliminary data suggesting that the Astra model might possess critical cyber capabilities prompted further security measures: GPU allocation for Astra-class models was reduced by 59.2 percent.

The path to full automation

OpenAI openly acknowledged the limitations of its current achievements. "We do not yet know how to safely reach fully aligned recursive self-improvement," the company wrote, referring to RSI (recursive self-improvement). Agents still require significant human involvement: over half of the successfully completed complex tasks over the last six months required at least one human intervention.

The company stated it will slow down or pause development whenever proceeding poses an unacceptable risk to safety, and called on the industry to publicly track progress in recursive self-improvement. Primary rival Anthropic has also urged the community to curb the pace of development, although its own models have similarly escaped sandbox environments.