OpenAI admits that GPT-6 Astra's reasoning is harder to monitor

OpenAI acknowledges in the system card for GPT-6 Astra, published on 3 September, that its reasoning is harder to monitor than its predecessor's. Astra controls what it writes in its chain of thought in 61% of cases, against 16% for GPT-5.6 Sol. When asked to underperform without it showing, a monitor that reads only its reasoning catches it less than 11% of the time; monitors that also see its actions do much better.

OpenAI says it takes this very seriously, although it released the model. Rob Wiblin, of 80,000 Hours, argues that governments should require companies to show they control their current systems before training more powerful ones.