GPT-6 Astra's System Card Reveals a Chain-of-Thought Controllability Jump From 16% to 61%

OpenAI's full system card for GPT-6 Astra adds hard benchmark numbers behind its safety overview, including external UK AISI and Apollo Research red-team results, and shows the model can now deliberately steer its own chain-of-thought far more than GPT-5.6 Sol could.