OpenAI AI Interns Cut Support Requests by 50%

Why this matters

OpenAI is telling us its own automation is working well enough to shut down human office hours, while its chief scientist says nobody, including OpenAI, has actually solved control of these systems. Those two facts sitting in the same release should give developers and founders pause rather than confidence. A drop in support tickets is a real, measurable signal, but it's an internal productivity metric, not evidence of safety or reliability at scale.

If your own team is racing to plug agents into debugging and model development workflows, Pachocki's warning is the more important data point here, not the efficiency numbers. Watch what OpenAI does next, not what it publishes: does it slow deployment pace to match its control research, or does the internal automation trend keep outrunning the safety work it admits is unfinished? For anyone building on top of these models, the practical takeaway is to treat "agents handle it now" claims as a productivity story, not a maturity story, and to ask labs directly what "solved control" would even look like before betting infrastructure on the assumption that it's close.