OpenAI's Astra model, expected this week, will use a reasoning technique called recurrent depth or opaque recurrence. Instead of thinking in a straight line from question to answer, the model loops through the same query multiple times in internal states before producing language.
Why safety researchers are worried
Normal reasoning models show their chain of thought: the sequential steps they took to solve a problem. That record is imperfect, but it lets humans and monitoring systems flag problems. Opaque recurrence leaves fewer readable traces. The model does more thinking in a form nobody outside can inspect line by line.
Redwood Research CEO Buck Shlegeris wrote he is extremely concerned, warning that if OpenAI scales up the technique, it could totally destroy chain-of-thought monitoring. AI safety researcher Zvi Mowshowitz said the approach risks a taboo OpenAI and Anthropic fought to establish: keeping reasoning visible for as long as possible.
OpenAI says monitoring will continue
OpenAI chief scientist Jakub Pachocki said the company has worked to preserve chain-of-thought monitoring since its first reasoning models and that it remains a core research goal. Astra's use of the technique is reportedly limited. But according to The Information, both Anthropic and Google are now studying recurrent depth, meaning the transparency tradeoff may spread across the industry.