According to CoinMeta, the Astra model that OpenAI is about to release adopts a cyclic reasoning architecture known as recurrent depth, also referred to as looped transformer. Unlike ordinary transformer models where each token generated fixedly passes through layers of the network, Astra allows the same piece of information to repeatedly go through the same set of layers and be calculated multiple times before output. This architecture was made public last year, and the Ouro model released by the Seed team also utilizes looped language model, enabling a group of transformer layers to operate in a cycle, thereby increasing the amount of reasoning and computation. As a result, smaller models can use more computational power to achieve results close to those of larger models. However, this architecture also brings security issues; some of the reasoning processes occur in internal, hidden states, making it impossible for humans to see the complete textual records and difficult to check whether the model is violating any rules. Therefore, OpenAI has limited the extent to which Astra uses recurrent depth in order to preserve a readable chain of thought, and additional monitoring measures are being prepared to be implemented.