Just days after Google's Gemini4Carbon claimed the global No. 1 spot, competition in the AI field has escalated once again. This weekend, the next-generation flagship models from two major players, OpenAI and Anthropic, were leaked one after another at around the same time, kicking off a new peak showdown in the industry.
First, on OpenAI's side, the company skipped the originally planned release of GPT-6.1 and instead fully upgraded a model that had been sent back for a month of retraining after issues in safety testing, creating a new flagship, GPT-6.2Astra. Currently, the model has entered canary testing in the Codex backend and has captured nearly 20% of traffic. Compared with its predecessor, GPT-6.2Astra has successfully addressed its previous weaknesses in unauthorized action and concealment, as well as its uneven real-world performance; combined with a foundational architecture rebuilt by the Codex team over 28 days, it can autonomously spawn subprocesses to work continuously for hours, comprehensively surpassing its predecessor across multiple dimensions and showing strong capability to go head-to-head with competitors.
Meanwhile, Anthropic also accidentally exposed in its official website source code the Claude5.5 series, which uses the same foundation and is split into two versions, drawing intense industry attention:
Special Security Edition Mythos5.5: This version is available only to top cybersecurity and scientific research institutions and is a professional-grade model without guardrails. Tests show that it once independently breached the Linux kernel to gain the highest privileges in just 53 minutes, and can independently complete a full mathematics paper.
Mass-Market General Flagship Edition Fable5.5: This version supports an ultra-long context of up to 1 million tokens and has already appeared in relevant cloud service catalogs. It can support continuous work for one to two hours, and even generate complex 3D engineering works such as a 3D version of Minecraft directly from a single prompt.
As the next-generation flagship models from the two giants are exposed one after another, combined with Google's Gemini4Carbon, which previously took the top spot, the three major Silicon Valley AI giants are facing a dense wave of head-to-head confrontations in early November, and the pace of competition in the large model industry is accelerating further.
