The Astra model, which OpenAI has not yet released, is reportedly equipped with advanced cybersecurity capabilities and uses a method known as “recurrent depth” or “looped transformer.” This technique allows information to be passed through the model’s internal layers multiple times, enabling more computation; however, a larger portion of the reasoning may remain outside the chain of thought expressed in natural language.
AI safety researchers say the approach, also known as “opaque recurrence,” could make models more difficult to monitor. Ryan Greenblatt said visible traces of thought could diminish and that the race for performance could push companies toward less observable architectures. Buck Shlegeris and Zvi Mowshowitz also voiced their concerns, while Mowshowitz suggested that regulation might be needed in the future.
OpenAI is keeping the technique’s use in Astra limited; it is said that a substantial portion of the chain of thought will remain readable. Jakub Pachocki said chain-of-thought tracking has been preserved and remains a core objective. Although Pachocki said Astra’s computational depth is in a range roughly twice that of GPT-4, OpenAI has neither confirmed nor denied the use of a looped transformer.
Why it matters
The discussion of this approach highlights the tension between greater computation in AI models and auditability. As the portion of reasoning that takes place outside the visible chain of thought grows, it may become more difficult for researchers to assess the model’s errors, safety issues and decision-making processes. Although it has been stated that OpenAI aims to preserve the readable chain of thought and limit the technique’s use, it remains unclear whether a looped transformer is used in Astra. This uncertainty raises questions not only about the model’s technical capacity but also about the extent to which safety oversight will be possible. Experts’ assessments regarding the possibility of regulation leave open the question of whether performance gains will come at the expense of less observable systems.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 12 news reports mentioning this name in the last 90 days; the most recent is dated 3 September 2026.