DeepSeek 2.0 moment? Zhipu’s market value tops 1 trillion, and GLM-5.2 dominates the screens

Author: Xu Chao, Wall Street Insights

On Monday this week, the total market capitalization of Hong Kong stocks in the Zhipu index broke through HKD 1 trillion during trading, with an increase of over 1900% this year. This is not just a rally of a single stock—the release of China's open-source large model GLM-5.2 is redefining the boundaries of global AI capabilities, and has also pushed the discussion of "DeepSeek 2.0" onto Wall Street trading floors.

In terms of performance, GLM-5.2 scored 74.4 on the FrontierSWE long-range programming benchmark, just about 1 percentage point below Anthropic's top model Opus 4.8 at 75.1, while surpassing GPT-5.5's 72.6, making it currently the highest-rated open-source weight model, priced approximately 72% to 82% lower than Opus 4.8.

Almost simultaneously, Anthropic was forced to shut down global access to its flagship models Fable 5 and Mythos 5—U.S. Commerce Department intervened citing export control regulations, requiring the company to obtain government licenses before providing related services to foreigners. The convergence of these two news items instantly shaped a narrative of "U.S. restrictions, China opening."

Contrasting sharply with the shock to DeepSeek early 2025, this round of capital did not withdraw from NVIDIA and U.S. AI stocks, but instead flowed into Chinese assets, showing characteristics of alternative trading rather than panic liquidation. The core question the market is re-pricing is: when high-performance open-source models can deliver capabilities close to closed-source models at less than one-tenth of the cost, combined with U.S. policies that directly cut off the global availability of closed-source models, has the competitive landscape of the AI industry undergone a structural shift?

GLM-5.2: Open Source Enters the Frontline Competition Radius of Closed Source for the First Time

The significance of GLM-5.2 lies in its pushing open-source models into a performance range previously dominated by closed-source labs.

According to data released by Zhipu, GLM-5.2 has 753 billion parameters, uses a Mixture of Experts (MoE) architecture, supports a 1 million token stable context window, and is fully open-sourced under the MIT license. On the FrontierSWE programming benchmark, GLM-5.2 scored 74.4, with Anthropic's Opus 4.8 at 75.1—about 1 percentage point difference—and surpassing GPT-5.5 at 72.6. In the PostTrainBench (testing agent training small models) benchmark, GLM-5.2 ranked second with 34.3 points, just behind Opus 4.8 at 37.2, and ahead of GPT-5.5 at 28.4.

Artificial Analysis rated GLM-5.2 at 51 points in its Intelligence Index v4.1, ahead of MiniMax-M3 (44), DeepSeek V4 Pro (44), and Kimi K2.6 (43), placing it between GPT-5.5 and Opus 4.8, making it the highest-ranked open-source model to date. Community researcher @jeremyphoward表示 commented that "GLM-5.2 is at least comparable to Opus 4.8 and GPT-5.5"; @matvelloso称其为 called it "the first open-source model to meet my daily use standards."

Gaps still remain. On the most challenging SWE-Marathon benchmark, GLM-5.2 scored 13.0, while Opus 4.8 scored 26.0; visual capabilities are also a current shortcoming. However, from an engineering deployment perspective, the IndexShare technology introduced by GLM-5.2—layer-crossing sparse attention top-k index reuse—significantly compresses inference computation for ultra-long contexts, making the cost of 1 million tokens much more feasible. AI research organization Proximal commented that GLM-5.2 is "the first model to truly narrow the huge technological gap between Anthropic/OpenAI and other model providers."

Pricing Logic: Frontline Capability Upgrades Still Support Premium

The pricing structure of GLM-5.2 offers a new reference framework for valuation at the AI model layer.

The input/output token prices of GLM-5.2 are about 72% to 82% lower than Opus 4.8. However, JPMorgan Chase's report notes that compared to GLM-5.1, GLM-5.2 is actually a price increase: GLM-5.1 used tiered billing, allowing some usage at lower rates; GLM-5.2 applies a uniform higher pricing tier, so the actual blended price paid by customers has increased. Since the performance improvements mainly come from reinforcement learning and post-training optimization rather than large-scale model expansion, the cost base remains relatively stable. This adjustment is expected to improve Z.ai's gross margins.

Based on this, JPMorgan Chase concluded: "Mature intelligent compression pricing, but GLM-5.2 shows that frontline upgrades can have the opposite effect." The bank believes that AI model pricing is showing a structural divergence: capabilities like basic dialogue, simple summarization, and standard code assistance—already commoditized—will continue to face price compression, with DeepSeek being a typical example; while frontier capabilities that unlock new workflows and improve task completion—especially in programming, agents, enterprise automation, and long-context scenarios—can still maintain or even increase prices under the logic of "paying for task completion rather than tokens."

For investors, this distinction has direct valuation implications: the monetization prospects of model-layer companies depend on their ability to continuously move toward more difficult, higher-value tasks, rather than merely scaling existing capabilities.

Anthropic Models Removed from Market: The Risk of Closed-Source Accessibility Turning from Concept to Reality

The sudden removal of Fable 5 and Mythos 5 turns the risk of closed-source commercial models' availability from an abstract discussion into a direct impact.

According to Bloomberg, Howard Lutnick cited Section 744.22(b) of the Export Administration Regulations, citing the "unacceptable risk" of these models being exploited by foreign military intelligence agencies, and demanded that Anthropic obtain a Commerce Department license before providing access to any foreign individuals worldwide, or face criminal and civil penalties.

A research report from Orient Securities cited media reports that Amazon researchers successfully bypassed security restrictions of Mythos models and found at least four security vulnerabilities in four different software under Fable 5 when prompted in specific ways. This was considered a key trigger for regulatory intervention. Anthropic then shut down global access to both models and publicly stated that government response was "disproportionate," warning that if the same standards are extended industry-wide, all new deployments of frontier models could effectively come to a halt.

According to Wall Street Insights, Anthropic's technical team held talks with officials at the U.S. Commerce Department this Monday.

Analysts believe that this incident impacts the industry chain on two levels: first, companies and developers relying on closed-source frontier models face business continuity risks, increasing demand for alternatives; second, open-source models with available weights and local deployment conditions have inherent advantages in controllability, and GLM-5.2 offers a performance close to the frontier at significantly lower costs at this critical juncture.

This regulatory trend has also attracted high attention from other AI labs. Insiders say OpenAI's Chief Strategy Officer Jason Kwon has informed employees that the company is assessing the impact of this policy trend, describing the current situation as "a rapidly evolving scenario with many unknowns." OpenAI's General Counsel Che Chang reminded staff internally that when facing regulatory uncertainty together, "they should not attempt coordinated responses, as antitrust rules apply here."

Market Outlook: Alternative Trading Rather Than Panic Liquidation, Computing Power Still Booming

This round of market activity differs fundamentally from the DeepSeek incident, but the medium- to long-term industry logic is being reassessed.

DeepSeek's shock was an unexpected black swan, directly triggering a sell-off in the U.S. AI sector. The release of GLM-5.2, however, was an event within highly anticipated expectations—the market has digested the Chinese open-source model competition for 18 months, and this validation is reflected in the re-pricing of Chinese local AI stocks, which are currently not experiencing systemic shocks. JPMorgan Chase characterized this market movement as "alternative trading" rather than "panic liquidation." After raising Zhipu's target price to HKD 1,800, the stock has risen to around HKD 2,400, surpassing even the latest target, indicating that market pricing has already moved ahead of analyst forecasts.

Orient Securities believes that many domestic models are leading the global performance rankings, most of which remain open-source; combined with the delisting of two top Anthropic models, the API call volume for domestic models is expected to further increase, and demand for computing power and token services based on domestic models is likely to maintain good growth and prosperity.

Rich Privorotsky also pointed out that the AI sector is currently facing a tug-of-war between two forces: on one side, accelerated application adoption and rising computing power demand; on the other, token deflation, questionable monetization prospects, and continuous stock supply expansion. The market is currently more focused on the latter. But from a medium- to long-term industry logic, falling costs and lower access barriers may simultaneously drive token consumption and computing power demand expansion. Analysts note that increasing open-source model share and high growth in computing demand are becoming core variables in the reevaluation of the AI industry chain.