OpenAI lowers ability to monitor new Astra model using opaque 'recurrent depth' method
OpenAI CEO Sam Altman acknowledged that monitoring its new Astra model decreased compared to the previous GPT-5.6 Sol version, citing a method called 'recurrent depth'.
OpenAI announced on Thursday that its ability to monitor the new Astra artificial intelligence model has decreased relative to the prior GPT‑5.6 Sol version. The company stated this capability reduction in an official blog post. A specific technique named 'recurrent depth' reportedly enables the system to process reasoning more efficiently by not revealing steps in human language. This approach contrasts with current models that display a chain of thought to show processing steps. OpenAI said monitoring Astra became harder, though it did not quantify the extent of the decrease or provide specific metrics regarding safety risks. Industry analysts view this move as prioritizing commercial speed over transparency, even as the firm maintains broad pledges on safety.
