OpenAI ties fading AI transparency to four training factors; GPT-6 Astra cited

OpenAI's latest analysis builds on earlier findings. As recently as this spring, the company considered it good news that models like GPT-5.4 could barely manipulate their inner monologue on command. That trend broke with GPT-6 Astra in September, when chief scientist Jakub Pachocki called its monitorability fragile and warned that no AI lab had solved monitoring well enough to keep scaling at full speed much longer.

The safety team systematically studied which training steps trigger this erosion. They traced it to four main causes across the training cycle: rising base-model capability, architecture choice (one design, "Architecture 2", showed a much steeper performance curve on silent reasoning tasks), reinforcement learning (which for Astra-class and new "Terra" models increased control over the chain of thought), and inheritance from teacher models during distillation.

The hunt for simple fixes was sobering. Since February 2026, OpenAI has automatically checked every training run to rule out direct penalties on the scratchpad, yet Astra still developed strong control over its scratchpad. Still, the data reveal a concrete tradeoff between safety, capability, and cost, and OpenAI says it urgently needs new monitoring methods beyond reading the scratchpad.