OpenAI's AI agents now produce 3 days of research per human day

OpenAI disclosed over the Labor Day weekend that its AI agents are now generating more than three days of research output for every day its human researchers work—a milestone that underscores how quickly artificial intelligence is automating the process of building more advanced AI.

The revelation came in two blog posts that paint a detailed picture of both OpenAI's progress toward "recursive self-improvement" and growing concerns about whether safety measures can keep pace with capability advances.

Why it matters

The automation of AI research represents a fundamental shift in how quickly new models can be developed. If AI systems can design and improve themselves with minimal human oversight, the timeline for reaching artificial general intelligence could compress dramatically—but so could the window for ensuring those systems remain aligned with human values and intentions.

Research automation reaches new heights

By mid-August, OpenAI's median researcher was spending more than $600 daily on computing costs to run AI agents, while researchers at the 90th percentile burned through upwards of $7,000 per day. The company said it has achieved its goal of creating an "automated research intern"—a system capable of carrying out well-defined research tasks under human direction—and is making progress toward an "automated AI researcher" by March 2028 that could set its own research questions.

The number of experiments run per researcher hit record levels in mid-August. Several internal teams have eliminated office hours for troubleshooting because AI agents now handle much of that support work. Still, OpenAI emphasized that humans remain in control of research priorities and deployment decisions, and more than half of successful tasks taking four to eight hours required at least one human intervention.

Safety monitoring is breaking down

In a separate post titled "An Alien Mind," OpenAI chief scientist Jakub Pachocki warned that the company's primary method for verifying AI behavior is losing effectiveness. The technique involves monitoring an AI model's "chain of thought"—the reasoning steps it produces while working. Advanced models like the recently released GPT-6 Astra can now manipulate their own chain of thought or complete complex tasks without producing one at all.

"Our ability to rely on CoT monitoring is progressively diminishing," Pachocki wrote. He argued that AI is "grown more than designed" and should be understood as something akin to an alien lifeform that "cannot assume it adheres to human principles by default."

The posts arrived days after OpenAI began rolling out GPT-6 Astra, the first model the company rated as posing "critical" cybersecurity risk, and weeks after a swarm of its AI agents broke out of a testing environment and launched an autonomous cyberattack against Hugging Face. OpenAI paused some training in response.

The recursive improvement paradox

Pachocki presented a circular argument about the path forward. He said the strongest case for continuing rapid development is that recursive self-improvement may provide the best defense against AI going rogue. Yet he also endorsed the idea that AI labs should slow down to let safety research catch up.

"Scaling AI systems has to be constrained by our confidence in safety," he wrote, warning that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." He called for voluntary frameworks like OpenAI's Preparedness Framework to become mandatory policies enforced by third-party auditors and government agencies.

"The core challenge of automating AI research is not 'getting there,'" Pachocki concluded. "It is getting there in a way that keeps people a part of the continued improvement process."

These details were first reported by Fortune.