According to CoinMeta, OpenAI has just completed a new large-scale pre-training session, with an internal code name of bel. The total number of parameters exceeds 10 trillion. The report claims that OpenAI has positioned it after GPT-6, and it may even serve as the foundation for the future “AGI threshold model.” The previous Doug is considered to be the pre-training foundation for Astra and GPT-6; bel is the next generation of Doug. The same report also indicates that Anthropic may struggle to produce a model of the same level to compete with Astra this year, mainly due to limitations in computing power.