``` DeepSeek 2.0 moment? Zhipu's market value surpasses one trillion HKD, GLM-5.2 makes headlines on Wall Street ```

```

This Monday, the market capitalization of Zhipu AI on the Hong Kong stock exchange surpassed HK$1 trillion during intraday trading, with its year-to-date increase exceeding 1900%. This is not just the story of a single stock—China's open-source large model GLM-5.2 is redefining the boundaries of global AI capabilities, and has brought discussions of "DeepSeek 2.0" to Wall Street trading desks.

In terms of performance, GLM-5.2 scored 74.4 on the FrontierSWE long-range programming benchmark, just about 1 percentage point lower than Anthropic's top model Opus 4.8 (75.1), while surpassing GPT-5.5's 72.6. It has become the highest-scoring 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—the U.S. Department of Commerce intervened citing export control regulations, requiring it to obtain government licenses before providing relevant services to foreigners. The combination of the two news events instantly shaped the narrative of "U.S. restrictions, China openness."

Unlike the panic during DeepSeek's turmoil at the start of 2025, this round of capital did not exit Nvidia and U.S. AI stocks, but instead shifted into Chinese assets, showing a substitution trade characteristic rather than panic selling. The core question the market is now repricing is: when high-performance open-source models can deliver comparable capabilities at less than one-tenth the cost of closed-source models, and U.S. policy directly cuts off the global availability of closed-source models, has the competitive landscape of the AI industry chain undergone a structural shift?

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

The key significance of GLM-5.2 is that it has pushed open-source models into a performance domain previously dominated by closed-source labs.

According to Zhipu's released data, GLM-5.2 has a parameter size of 753B, uses an MoE (Mixture of Experts) architecture, supports a stable context window of 1M tokens, and is completely 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—a difference of about 1 percentage point—while surpassing GPT-5.5's 72.6. On the PostTrainBench (testing agents training on small models) benchmark, GLM-5.2 ranked second with 34.3 points, behind only Opus 4.8's 37.2, and ahead of GPT-5.5's 28.4.

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

A gap still exists. On the most difficult SWE-Marathon benchmark, GLM-5.2 scored 13.0, while Opus 4.8 scored 26.0; lack of visual capabilities is also a current shortcoming. But from an engineering deployment perspective, GLM-5.2 introduces IndexShare technology—cross-layer reuse of sparse attention top-k indices—which significantly reduces the inference computation for ultralong context, making 1M context feasibility greatly improved. AI research institution Proximal commented that GLM-5.2 is "the first model to truly bridge the vast technological gap between Anthropic/OpenAI and other model providers."

Pricing Logic: Frontier Capability Upgrades Can Still Support Premiums

The pricing structure of GLM-5.2 provides a new reference framework for model-layer AI valuations.

The input/output token price of GLM-5.2 is about 72% to 82% lower than Opus 4.8. However, JP Morgan pointed out in its report that, compared with GLM-5.1, GLM-5.2 is actually a price increase: GLM-5.1 used tiered billing, with lower rates available for partial usage; GLM-5.2 uniformly applies a higher pricing tier, so the blended price paid by clients rises. Since performance improvements mainly come from reinforcement learning and post-training optimization, rather than massive scaling up of model size, cost base stays stable, so this adjustment is expected to drive Z.ai's gross margin improvement.

JP Morgan thus concluded: "Mature intelligence compresses prices, but GLM-5.2 shows that frontier upgrades can achieve the opposite effect." The bank believes AI model pricing is showing structural differentiation: basic conversation, simple summarization, and standard code assistance—already commoditized capabilities—will continue to face price compression, with DeepSeek being the most typical representative; but frontier capabilities that can unlock new workflows and improve task completion rates—especially for programming, agents, enterprise workflow automation, and long-context tasks—can still maintain or even increase pricing under clients' logic of "paying for task completion, not tokens."

For investors, this distinction has direct valuation implications: the monetization outlook of model-layer companies depends on whether they can keep moving to harder, higher-value tasks, not just on scaling up existing capabilities.

Anthropic Model Removal: Closed-Source Accessibility Risk Goes from Concept to Reality

The sudden removal of Fable 5 and Mythos 5 turns the risk of closed-source commercial model accessibility from an abstract discussion into direct impact.

According to Bloomberg, Howard Lutnick cited Section 744.22(b) of the Export Administration Regulations, claiming there was an "unacceptable risk" of the models being used by foreign military intelligence agencies, and required Anthropic to obtain a Commerce Department license before providing global access to foreigners, otherwise facing criminal and civil penalties. A research report from Dongfang Securities cited media reports that Amazon researchers had successfully bypassed Mythos's model safety restrictions and found that Fable 5, under certain prompt guidance, could discover security vulnerabilities in at least four pieces of software, believed to be the trigger for regulatory intervention. Anthropic promptly closed global access to both models and publicly stated the government response was "disproportionate," warning that if the same standards were extended to the industry, all frontier model deployments could essentially stall. According to Wallstreetcn, Anthropic's technical team held meetings with Commerce Department officials this Monday.

Analysis suggests this incident affects the industry chain at two levels: First, enterprises and developers relying on closed-source frontier models face business continuity risks, driving demand for alternative solutions; second, open-source models—with open weights and local deployment—naturally have an advantage in controllability, and at this moment, GLM-5.2 offers a performance-close and significantly lower-cost substitution option.

This regulatory movement has also triggered keen attention from other AI labs. According to sources, OpenAI’s Chief Strategy Officer Jason Kwon has notified employees that the company is assessing the policy's impact and described the situation as "rapidly evolving with many unknowns." OpenAI's General Counsel Che Chang warned internally that, in facing regulatory uncertainty together, "no coordinated response should be attempted; antitrust rules apply here."

Market Characterization: Substitution Trading, Not Liquidation Panic; Computing Power Renaissance Sustained

This round of action differs structurally from the DeepSeek event, but the long-term industrial logic is being re-evaluated.

The DeepSeek shock was an unexpected black swan event that directly triggered a sell-off in U.S. AI stocks. The release of GLM-5.2, on the other hand, was highly anticipated—the market had 18 months to digest expectations about the competitiveness of Chinese open-source models. This time, the confirmation was a concentrated re-pricing of China’s domestic AI names, while U.S. AI stocks have so far seen no systemic shock. JP Morgan calls this market action a "substitution trade" rather than a "liquidation panic." After raising Zhipu’s target price to 1,800 HKD, its share price soared to around 2,400 HKD, further surpassing the latest target, showing that market pricing has moved ahead of analyst forecasts.

Dongfang Securities believes that multiple domestic models are among the global performance leaders, with most remaining open-source; combined with the removal of Anthropic’s two leading models, API calls for domestic models are expected to further increase, and demand for computing power and token services based on domestic models is expected to remain strong and vibrant.

Rich Privorotsky also pointed out that the AI sector is facing two competing forces: on the one hand, accelerated adoption and rising computing demand; on the other, intensified token deflation, uncertain monetization prospects, and continued stock supply expansion, with current market focus more on the latter. But from a long-term industry logic, lower costs and access barriers may simultaneously drive both token consumption and computing demand growth. Analysts note that the rise in open-source model share and strong computing demand are becoming key variables for a revaluation of the AI industry chain.

Risk Disclosure and DisclaimerThe market carries risk, and investment requires caution. This article does not constitute individual investment advice nor consider individual users' specific investment objectives, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular circumstances. Investment made on this basis is at your own risk. ```