Daily Tokens
November 12, 2026

AI · November 12, 2026 · Moderate impact

OpenAI research traces why frontier model reasoning is becoming less transparent

OpenAI's safety team systematically investigated why the chains of thought in frontier models such as GPT-6 Astra are increasingly evading oversight. The analysis identifies four training steps that drive the erosion, including architecture choices and reinforcement learning. The team concludes the loss of transparency is not an unavoidable law but also reflects tradeoffs between capability and compute cost.

It suggests that keeping model reasoning inspectable is a design choice that teams can trade off against capability and cost, not something that is fixed.

Go deeper

More from AI on Thursday, November 12