Sources
AI Explained's Astra review centres on chain-of-thought monitorability slipping, not the benchmark wins
Published 2026-09-04 (29 minutes), the video's framing is that Astra arrived on a day benchmark-makers got humbled and AI safety researchers got unnerved, with a dedicated segment on monitorability losing hold of Astra's chains of thought and a closing 'CoT Control' section. It cites OpenAI's own deployment safety PDF at deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf alongside the launch post. For builders relying on reading a model's reasoning trace to audit agent behaviour, a frontier model whose CoT is drifting out of legibility is the operationally relevant part of this launch.
↳ Follow the thread