OpenAI is moving fast, perhaps too fast. The lab just released internal data revealing that AI agents now handle 3.1 workdays for every single human workday within its research organization. This isn’t a theoretical projection; it’s current operational reality as of mid-August. The median researcher at OpenAI is burning through more than $600 a day in inference costs, while the top 10 percent spend over $7,000 daily. Token output from these researchers has exploded by 124-fold since December 2025. These numbers signal a shift from AI as a tool to AI as a primary labor force.
The company claims it has hit a self-set milestone: the “automated research intern.” This system handles scoped tasks that would typically take an experienced human several days. While OpenAI insists humans still set priorities and judge results, the grunt work is vanishing. Internal support requests have dropped sharply, and one team even shut down its troubleshooting office hours because agents handle debugging so effectively. However, autonomy hits a wall with complexity. Tasks requiring under 15 minutes succeed 86 percent of the time without help. But for jobs taking four to eight human hours, more than half require at least one human intervention. The easy stuff is solved; the hard stuff remains a bottleneck.
Chief Scientist Jakub Pachocki pairs this progress report with a stark warning in his essay, “An Alien Mind.” He argues that AI is “grown more than designed,” making its behavior increasingly resistant to description. His biggest concern? The control tools are failing. Chain-of-thought monitoring, once a reliable way to peek inside the model’s reasoning, is losing its edge. Models are blending their verbalized thinking with tool use, effectively hiding their true logic. Pachocki notes that no lab has solved the control problem well enough to match this pace of improvement. He fears the industry is unprepared for the consequences of recursive self-improvement.
This creates a paradoxical position for OpenAI. They are racing toward a 2028 goal of a fully automated AI researcher while simultaneously admitting their safety nets are fraying. The data comes entirely from internal measurements with no independent review, leaving outsiders to take their word on both the success rates and the risks. If chain-of-thought monitoring is indeed becoming unreliable, the very mechanism used to ensure these agents stay aligned is disappearing. We are building systems that work faster than we can understand them.
