OpenAI Completes Codenamed 'Bel' Pretraining, Potential Foundation for GPT-6 and Astra

According to a disclosure by X user synthwavedd, OpenAI recently completed the next phase of pretraining, internally codenamed "Bel," which is the successor model to the previously named "Doug." The model has a total parameter count exceeding 10T, with a scale similar to GPT-4.5, and is expected to become the foundation model for Astra and GPT-6 after further reinforcement learning, potentially serving as the base for the post-GPT-6 era and even models approaching the AGI threshold. The news indicates that this development has raised concerns at Anthropic. OpenAI believes Anthropic currently lacks an effective strategy to counter Astra's public release; sources familiar with the matter say this assessment is largely accurate, primarily due to computational constraints. Anthropic is preparing for the possibility that OpenAI may dominate for the remainder of this year, but expects to regain its lead early next year.