GPT-6 Astra's looped reasoning means its visible chain of thought is narration, not a trace

GPT-6 Astra uses looped transformer blocks that run roughly 44 passes over the same weights, meaning the actual computation is hidden inside recurrent cycles users cannot inspect. The reasoning text the model produces is written after the fact as a plausible summary, not a real-time record of how the output was derived. This creates a blind spot for agent evaluation and code review workflows that treat chain-of-thought transcripts as reliable evidence of the model's thinking. Evaluators are advised to instead instrument observable outputs such as tool calls, file changes, and final diffs, which can be verified independently. Additionally, forcing a model to show its reasoning alters how it allocates compute passes, meaning benchmarks run on transparent variants may not reflect the behavior of the default deployment.