Jefferies Evaluates Zhipu GLM-5.2: Enters Global Top Tier, But Hard to Replace Anthropic

Huo Xing Cai Jing reports that on June 22, Jefferies stated that Zhipu’s GLM-5.2 has entered the top three globally among large models, marking a significant milestone in China’s AI development; however, U.S. restrictions on access to Anthropic’s top models are unlikely to generate substantial revenue growth for Zhipu. In a China technology sector report released on June 17, Jefferies noted that GLM-5.2 ranks third in the Global Model Intelligence Index, behind only Anthropic and OpenAI—marking the first time a Chinese model has entered the global top three. The report highlights that GLM-5.2 excels in programming and long-term agent workflows and has become one of China’s leading large models. Zhipu released GLM-5.2 via its Coding Plan on June 13 and opened its API on June 17. The model employs a 744-billion-parameter MoE architecture with approximately 40 billion active parameters, supports a 1-million-token context window, and releases its weights under the MIT license. According to Jefferies, in Artificial Analysis’s Model Intelligence Index, GLM-5.2 ranks third globally and first in China; in coding and agentic metrics, it ranks fourth and second globally, respectively, and first in China. On Arena.ai’s frontend programming leaderboard, GLM-5.2 Max ranks second, behind Fable 5 High and ahead of Claude Opus 4.7 and 4.8. The report also notes that U.S. restrictions on access to Anthropic’s Claude Mythos 5 and Fable 5 may be difficult to sustain long-term. After Anthropic released these models on June 9, the U.S. government, citing national security concerns, required it to restrict access by foreigners—including non-U.S. citizens working within the U.S. Due to the short-term difficulty of distinguishing user identities, Anthropic temporarily suspended global access and strengthened identity verification. Jefferies believes that this “U.S.-only access to frontier models” arrangement is operationally challenging and could harm the U.S. AI ecosystem. A large number of U.S.-based AI researchers and engineers are foreign-born or non-U.S. citizens; excluding them from frontier models could slow research progress and even drive talent and demand overseas. However, the report concludes that restrictions on Fable 5 are unlikely to translate into significant revenue gains for Chinese model companies. Jefferies’ channel surveys indicate that most users are not directly switching from Claude to GLM-5.2. GLM-5.2’s programming capabilities are considered comparable to Claude Opus 4.7 but still lag behind Opus 4.8. Even if developers adopt Chinese open-source models, they may utilize them via local deployment, cloud providers, or inference platforms like OpenRouter—without directly contributing revenue to model developers. The report further states that developers typically use multiple models for different tasks—including code generation, debugging, and long-context tasks—rather than fully migrating to a single model. This limits GLM-5.2’s ability to monetize potential traffic spillover from Anthropic’s restrictions. Jefferies points out that Zhipu still faces insufficient high-end inference computing power, particularly for long-context and agent workflows, which consume substantial resources and may hinder enterprise-grade delivery. The report also notes that competition in China’s large model pricing market is intense: although GLM-5.2’s performance is approximately 26% higher than GLM-5.1, its API pricing remains unchanged, indicating that model providers lack significant pricing power. Additionally, U.S. model vendors’ anti-distillation technologies and export controls pose long-term risks. Jefferies states these measures may impede Chinese models’ ability to catch up with frontier capabilities and could re-expand the performance gap between U.S. and Chinese models. The report’s core conclusion is that GLM-5.2 demonstrates China’s models have entered the global top tier—but they have not yet truly “worn Anthropic’s shoes.” Under constraints including revenue conversion, computing supply, price competition, and U.S. technological restrictions, Zhipu is unlikely to achieve meaningful commercial gains from Anthropic’s access limitations in the near term.