OpenAI is stepping up the research and development of the next generation flagship model. Circular reasoning and security alignment are key. OpenAI is stepping up the research and development of the next generation flagship model. However, due to the impact of security alignment and test infrastructure construction, there is still uncertainty about the release time.

OpenAI is stepping up the research and development of the next generation flagship model. Circular reasoning and security alignment are key. OpenAI is stepping up the research and development of the next generation flagship model. However, due to the impact of security alignment and test infrastructure construction, there is still uncertainty about the release time. According to information disclosed by "The Information" in October, OpenAI employees expect that the company will release one or two new models before the end of 2026, but it is not yet certain whether the flagship upgraded model under development will be among them. It is reported that the new flagship model uses the same Recurrent Transformer technology as Astra, namely "Recurrent Depth". This technology allows the model to process the current word unit through the same Transformer layer multiple times before generating the next word unit, which can improve mathematical reasoning and programming capabilities and enable smaller models to achieve stronger performance at lower costs. However, circular reasoning also shifts part of the computational process to internal numerical representations that are difficult to review by humans, reducing the ability to monitor model behavior through readable thought chains. According to reports, OpenAI has limited the number of Astra cycles to maintain the readability of the inference process, and has shifted about 10% of the total computing power to algorithm alignment and security research, with the relevant computing power scale reportedly reaching hundreds of megawatts. A GPT-6.1Sol technical article previously published by Microsoft accidentally disclosed the details of circular reasoning and subsequently deleted the relevant content. OpenAI also needs to improve its security testing and, in accordance with relevant arrangements, undergo a security assessment by the U.S. government before the model is delivered to external partners and before it is officially released to the public. This research and development progress shows that the balance between inference performance improvement and security monitorability is becoming an important challenge for the release of next-generation large models. via AI News (author: AI Base)