OpenAI completes Bel pretrain with 10T+ parameters, targeting Astra and GPT-6
On August 25, OpenAI completed a major new pretrain codenamed Bel, the successor to Doug, featuring over 10 trillion total parameters similar in scale to GPT-4.5. Bel is expected to serve as the base model for Astra and GPT-6 after further reinforcement learning.
According to X posts on August 25, OpenAI has completed a major new pretrain codenamed Bel, the successor to Doug. Bel features over 10 trillion total parameters, similar in scale to GPT-4.5. Bel is expected to serve as the base model for Astra and GPT-6 after further reinforcement learning, and potentially even an AGI-threshold model.
According to reports, this development has Anthropic concerned. The company reportedly has no strong public response to Astra ready, largely due to compute constraints, and expects OpenAI to dominate for most of the year.
The information originates from X platform leaker leo and has not been officially confirmed by OpenAI.