Several newly developed models from OpenAI are currently garnering significant attention. Among them, the most powerful model, Astra, is anticipated to be released as early as next Thursday. At present, Astra has broadened its testing reach and assumed control over the foundational code for OpenAI’s self-developed chip, named ‘Jalapeño’. When it comes to writing core code, Astra outpaces top human engineers by 1.8 times, and there are already indications of internal AI self-recursion. Furthermore, OpenAI has finalized the development of the ultra-large-scale pre-trained model, ‘Bel’. This model boasts over 10 trillion parameters and is poised to serve as the next-generation cornerstone for the ‘post-GPT-6 era’, with aspirations to achieve Artificial General Intelligence (AGI). The ‘Bel’ model outperforms Astra in numerous aspects and has the capability for continuous evolution, earning it the moniker of the ‘killer’ of Claude Fable. OpenAI is confident that, leveraging its computational power advantage, the company is set to sustain its leading position from the latter half of 2026 through 2027. In contrast, Anthropic may find it challenging to keep pace due to a scarcity of computational resources. OpenAI executives have also unveiled the future roadmap for the application layer. This includes recursive self-improvement aimed at achieving a commercial closed loop, the introduction of an Ultra Fast mode that will revolutionize interaction methods, a transition in computing paradigms towards large-scale cloud-based Agent clusters, and the integration of ChatGPT and Codex into a ‘Personal AGI’.