OpenAI's GPT-6 Astra Is Harder to Monitor After Reports of Recurrent-Depth Reasoning

OpenAI has begun rolling out GPT-6 Astra, a frontier model whose written reasoning is harder for researchers to inspect even as the company reports large gains in coding, computer use, and cybersecurity. Reporting ahead of the launch said Astra uses a limited form of recurrent depth, also called opaque recurrence: the model loops the same query through shared Transformer layers instead of spelling out every intermediate step, which can lift math and coding performance and cut costs while leaving fewer readable traces of potentially dangerous behavior.