OpenAI is accelerating development of its next-generation flagship model, but the release timing remains uncertain due to safety alignment and testing infrastructure considerations. According to information disclosed by The Information in October, OpenAI employees expect the company to release one or two new models by the end of 2026, but it remains unclear whether the flagship upgrade model under development will be among them.
It is reported that the new flagship model uses the same recurrent Transformer technology as Astra, known as "Recurrent Depth." This technology allows the model to pass the current token through the same Transformer layers multiple times before generating the next token, which can improve mathematical reasoning and programming capabilities, and enable smaller models to achieve stronger performance at lower cost.
However, recursive reasoning also shifts part of the computation process into internal numerical representations that are difficult for humans to audit, weakening the ability to monitor model behavior through readable chains of thought. It is reported that OpenAI has limited Astra's number of loops in order to maintain the readability of the reasoning process, and has redirected about 10% of its total computing power toward algorithmic alignment and safety research, with that computing capacity reportedly amounting to hundreds of megawatts.
A technical article on GPT-6.1Sol previously published by Microsoft unexpectedly disclosed details of cyclical reasoning, and the relevant content was subsequently deleted. OpenAI still needs to complete safety testing and, in accordance with the relevant arrangements, undergo safety assessments by the U.S. government both before the model is delivered to external partners and before its official public release. This development progress indicates that the balance between improved reasoning performance and monitorable safety is becoming a major challenge for the release of next-generation large models.