OpenAI reportedly pre-trained the Bel model, aiming for the AGI threshold According to a leak from @synthwavedd, OpenAI has recently completed pre-training its next-generation model, internally codenamed 'Bel.' This model succeeds the previous 'Doug' model and boasts over 10 trillion parameters, comparable in scale to GPT-4.5. It’s said that OpenAI aims to make Bel the foundational model after GPT-6 (combined with further reinforcement learning), potentially pushing toward the AGI (Artificial General Intelligence) threshold. The leak also mentions that this progress has raised concerns at Anthropic. Reportedly, OpenAI believes Anthropic is currently unprepared to compete with Astra (rumored to be the codename for OpenAI’s next-gen product) for a public release, mainly due to compute limitations. Anthropic is likely to remain on the back foot for the rest of this year but is expected to catch up by early next year.