OpenAI completes 'Bel' pre-training, which may serve as the foundation for GPT-6 and Astra

Huo Xing Finance reports that, according to X user synthwavedd, OpenAI has recently completed its next-stage pre-training, internally codenamed "Bel," succeeding the prior model "Doug." The model exceeds 10 trillion parameters in scale, comparable to GPT-4.5, and is expected, after further reinforcement learning, to serve as the foundational model for Astra and GPT-6, potentially forming the base for models beyond GPT-6 approaching the threshold of AGI. The report indicates that this development has raised concerns at Anthropic. OpenAI believes Anthropic currently lacks an effective response to the public release of Astra; insiders confirm this assessment is largely accurate, primarily due to computational constraints. Anthropic is preparing for the possibility that OpenAI will dominate for the remainder of this year, but expects to regain leadership by early next year.