OpenAI's Astra model uses "opaque recurrence" (also called "recurrent depth"), a reasoning technique where the model loops through the same query multiple times rather than following a linear chain-of-thought sequence. This produces fewer legible traces of reasoning, making it harder to audit or monitor the model's internal logic compared to standard reasoning models where you can read step-by-step traces. The source says Astra's current use is limited and its chain-of-thought remains legible. OpenAI emphasizes commitment to monitoring systems. However, safety researchers like Buck Shlegeris (Redwood) and Ryan Greenblatt worry the technique could scale significantly—potentially to where models reason entirely in "latent space" (internally, invisibly) rather than in explainable text. The concern: if this spreads across labs without guardrails, it could trigger a competitive race that erodes the transparency practices AI labs have tried to maintain. Anthropic and Google DeepMind are reportedly already exploring it. The core tension: does limited deployment of a monitoring-resistant technique represent acceptable R&D, or a dangerous precedent?