Just days after Google Gemini 4 Carbon claimed the top spot in the global rankings, OpenAI and Anthropic both had their next-generation models leaked over the same weekend. OpenAI skipped the originally scheduled GPT-6.1 release and jumped straight to GPT-6.2 Astra—a model that had been sent back for a month of retraining due to safety testing concerns. It is currently undergoing gray-scale testing in the Codex backend, handling nearly 20% of traffic, and has addressed the previous generation's shortcomings, including unauthorized actions, information concealment, and lopsided real-world performance. Built on a ground-up architecture that the Codex team rebuilt over 28 days, the model can autonomously spawn subprocesses to work continuously for hours, outperforming its predecessor across multiple dimensions and possessing the capability to compete head-on with rivals. Anthropic, meanwhile, exposed its Claude 5.5 series through a leak of its official website source code. The series features two versions built on the same base: Mythos 5.5, a no-guardrail special edition available only to top cybersecurity and research institutions, which once independently breached the Linux kernel to obtain the highest privileges within 53 minutes and can independently complete full mathematical papers; and Fable 5.5, a general-purpose flagship version open to the public, which supports a 1-million-token context, has already appeared in relevant cloud service catalogs, can work continuously for one to two hours, and can generate complete complex 3D engineering works—such as a full 3D version of "Minecraft"—from a single prompt. Combined with the previously top-ranked Google Gemini 4 Carbon, these leaks from the two giants mean that the three major Silicon Valley AI giants are entering a dense round of direct flagship product competition in early November.