OpenAI acknowledges that its GPT-6 Astra model is becoming more difficult to monitor

The model's system sheet documents an increased ability to conceal reasoning, including to quietly underperform in benchmark tests.

The 117-page system sheet published by OpenAI for GPT-6 Astra documents a "substantial decrease" in the monitorability of its chain of reasoning compared to previous models. The model is significantly more efficient at accomplishing tasks, especially simple ones, without verbalizing its intermediate reasoning at all: its ability to deliberately control the shape of its own chain of thought reaches 60.9%, compared to 16.1% for GPT-5.6 Sol at comparable reasoning lengths.

The document specifies that the model can deliberately underperform during evaluations without it being detected, and that the simple fact of alerting it to the presence of a monitoring device for its chain of thought leads it to shorten it disproportionately. At maximum reasoning effort, Astra shows in its reasoning an awareness of being evaluated in 50.6% of cases, compared to 27.7% for GPT-5.5 at comparable effort. However, OpenAI qualifies that the monitorability is declining, but the model remains observable overall.