When will the large-scale models in China reach the level of Fable? Musk predicts by the first quarter of next year, Tang Jie says it won't take that long.

This week, a major news in the global AI field is undoubtedly the release and open-sourcing of the new flagship large model GLM-5.2 by KNOWLEDGE ATLAS on Wednesday. The model adopts the MoE sparse mixture expert architecture, with a total of 744B parameters and around 40B activation parameters, supporting long context processing of 1M tokens. On the Artificial Analysis comprehensive ranking list, GLM-5.2 achieved a high score of 51 points, ranking at the forefront among all available models, on par with Claude Opus 4.8, and also ranked first among global open-source models. With the release of GLM-5.2, coupled with the earlier export control ban on Anthropic's advanced models by the United States, a public debate on the "race against time for large models between China and the US" has ensued on overseas social media platforms, even attracting the participation of the world's richest person Musk and KNOWLEDGE ATLAS's founder Tang Jie. On June 18th, a user on the X platform posed a core question to independent AI researcher Teortaxes, "When will China's large models reach the level of capability of Anthropic's 'Fable'?" Fable 5 is the strongest model Anthropic has opened to the public so far, attracting attention for its powerful long task processing and code engineering capabilities. It belongs to the Mythos model's sub-model and was previously only open to a few security organizations, critical infrastructure companies, and trusted researchers. The model has extremely strong capabilities in network security and can help defenders find software vulnerabilities, but it may also help attackers reduce attack costs. Earlier this month, the US government, citing national security reasons, requested to halt access to Fable 5 and Mythos 5 for foreign entities. In response, Teortaxes gave an answer of 7 months. Teortaxes believes that, excluding visual functions, the overall capabilities of GLM-5.2 are in the range of Opus 4.7 to 4.8. Mythos reached the Preview state (i.e., functionality Opus 4.8) in early February 2026, and extrapolating, the time window for the Chinese model to reach the complete Fable level may be around November to December of this year. In comparison, Musk's prediction is relatively more conservative. He gave a timeframe of probably the first quarter of next year (Q1). In a reply, Musk emphasized that Chinese large models can easily catch up in benchmark tests, but if measured in "true usefulness," reaching the level of Q1 next year would already be quite impressive. Interestingly, as the central figure in this discussion, KNOWLEDGE ATLAS founder Tang Jie's response was more concise and direct: "won't take that long." This statement implies a strong confidence from the top AI teams in China in their own technological iteration speed. According to KNOWLEDGE ATLAS, as introduced earlier this week to the "Science and Technology Innovation Board Daily" reporter, GLM-5.2 is the company's most powerful open-source model to date, focusing on two core aspects: converting 1M token context from "on-paper parameters" to "production-ready" and pushing the long-range coding capabilities further. Zhang Youyu, the Secretary-General of the AI Special Committee of the Beijing Computer Society and a specially-appointed researcher at Peking University, told the reporter from the "Science and Technology Innovation Board Daily" that the release of KNOWLEDGE ATLAS GLM-5.2 marks a critical breakthrough for domestic models in programming niche scenarios. Firstly, breaking the duopoly, relying on the top global programming performance and high cost-effectiveness, GLM-5.2 has significant advantages in the long-context programming scenario, reshaping the industry landscape and giving rise to the "new triumvirate" of KNOWLEDGE ATLAS, OpenAI, and Anthropic. Secondly, although it has not achieved full-dimensional surpassing, in the current majority of medium-to-high frequency development scenarios, the model can already fully serve as an alternative to top overseas models. "However, the core weaknesses still exist, in the deep mathematical logic reasoning and the complex integration of cross-domain knowledge, GLM-5.2 still has a certain technical gap compared to the top overseas levels, which is the next area that needs to be focused on." Astounding low cost Despite the existence of some performance drawbacks, the advantage of Chinese AI large models in another field has increasingly attracted the attention of the global AI industry - cost. A user named Hassan on the X platform expressed on Thursday that the design of the GLM-5.2 model is simply crazy. "I asked GLM 5.2 (left) and Opus 4.8 (right) to help me design a landing page, and you can't even tell the difference between the two. GLM costs 0.06 dollars, while Opus costs as much as 0.49 dollars. The cost is more than 6 times lower, while being faster and saving more tokens. This is another victory for open-source AI." Financial Union Society reported last week that facing the rapid rise in the cost of artificial intelligence, large enterprises and startups around the world are increasingly turning to lower-priced AI model tools, including many from China. This shift is directly pushing significant price pressure onto the leading US industries OpenAI and Anthropic. According to startup company Vercel, the share of AI usage by DeepSeek on their platform has soared from a meager 1% in April to 17% in May. On another startup company OpenRouter platform handling AI queries, since mid-May, DeepSeek has firmly occupied the throne of the most commonly used AI model vendor. In response to this, the well-known financial blog website zerohedge stated that this may be the most important chart in the world: the arms race of artificial intelligence. The question is, when the industry realizes that the "10% intelligence gap" between the large models of China and the US may correspond to a "90% cost advantage," global capital will also undergo a profound reflection: has the over 5 trillion dollars in capital allocation to the AI field in the past few years been severely misallocated? In any case, when the competition of large models shifts from simply "parameters and high walls" to "utility, open-source, and cost-effectiveness," the Chinese power represented by KNOWLEDGE ATLAS GLM-5.2, is rewriting the rules of global AI competition with an iteration speed that exceeds external expectations. This article is reproduced from the "Financial Union Society" APP, written by Xiao Xiang; GMTEight Editor: Song Zhiying.