Open Source vs. Closed Source: How Zhipu Broke Through the Global AI 'Blockade'

On June 22, 2025, Zhipu (02513.HK) rose 15.09%, reaching a market cap of over HKD 1.07 trillion. Just one day later, on June 23, the stock fell 9.96%, with market cap dropping to HKD 967.5 billion. In just half a year, this Chinese AI company jumped from tens of billions to trillions in market cap. The market even compares it to 'three Meituans, four JDs' to describe the shock from AI.

Behind this dramatic price swing is a profound route battle in the global AI industry: open source vs. closed source, two fundamentally different development models colliding.

Targeting Fable 5: A Precise 'Timing' Game

Anthropic has long been one of the most stringent US large-model companies regarding regional access control. In September 2025, it further tightened service terms, upgrading restrictions from geographic blockade to ownership blockade; in April 2026, it introduced identity verification, requiring some users to submit government-issued ID.

In June, the blockade escalated further. On June 9 local time, Anthropic released its strongest models ever, Claude Fable 5 and Claude Mythos 5. But just three days later, the US Department of Commerce banned all foreign nationals from accessing these two models, including foreign employees of the company.

Right after the ban, Zhipu responded almost instantly, publishing an article titled 'To Developers: GLM-5.2 Fully Open, Frontier Intelligence Belongs to Everyone,' stating clearly, 'In a moment when some frontier models suddenly become unavailable, we choose to believe in another path: frontier intelligence should not belong to only a few, nor be taken back at any time by a few rules.' On June 17, Zhipu officially launched and open-sourced GLM-5.2 under the most permissive MIT license, allowing free commercial use globally.

This is not the first time Zhipu has responded to Anthropic's blockade. In September 2025, Zhipu quickly launched a 'Claude API User Special Relocation Plan' in response to Anthropic's tightened policies.

Cathedral and Bazaar: The Underlying Logic of Two Routes

A 'bible' circulates among programmers—'The Cathedral and the Bazaar'—metaphorizing two development models: the cathedral is built by a dedicated team in a closed environment, while the bazaar relies on open collaboration on the internet. The book proposed the famous Linus's Law: 'Given enough eyeballs, all bugs are shallow.'

Anthropic chose the 'Cathedral' model, using strict closed-source strategy to maintain technical advantage and business barriers; Zhipu takes the 'Bazaar' route, leveraging open-source strategy to engage the global developer ecosystem. When Anthropic locks the cathedral door, Zhipu's bazaar paves the road to everyone's feet.

According to release data, GLM-5.2 focuses on long-horizon tasks, enabling 1M lossless context; within 1M context length, it reduces unit token computation by about 65.5%; under similar token budgets, its coding ability lies roughly between Claude Opus 4.7 and Opus 4.8. On the long-horizon coding benchmark FrontierSWE, GLM-5.2 scored 74.4, only about 1 percentage point behind Anthropic's top closed-source model Opus 4.8, and surpassing GPT-5.5.

AI research institution Proximal assessed that GLM-5.2 is 'the first model to truly narrow the huge technical gap between Anthropic/OpenAI and other model providers, and the strongest open-source model to date.'

The Rise of the Open-Source Route and Geopolitical Games

Since DeepSeek's emergence set off an open-source wave, China's large-model industry has steadily formed a competitive path combining 'capability pursuit + open-source ecosystem.' After a year and a half of accumulation, on the world's largest model aggregation platform OpenRouter, Chinese model call volume has risen from about 1.2% at the end of 2024 to over 50% now. Last week, models from DeepSeek, Xiaomi, MiniMax, and Tencent ranked top four in usage.

Dongfang Securities believes open-source models, with their open weights, independent controllability, and local deployability, become a better solution to mitigate geopolitical risks and ensure business continuity. In the future, downstream users may no longer solely pursue model performance but increasingly turn to models that are stably available, independently controllable, and sustainably accessible. This trend will directly boost open-source ecosystem market share.

An interesting detail: shortly after Zhipu released GLM-5.2, an X user asked an independent researcher 'when Chinese large models will reach Fable-level capability.' Teortaxes estimated a seven-month gap, and Musk replied that it might be near by Q1 2027. In response, Zhipu CEO Tang Jie directly said, 'Not that long.'

AGI Checkpoint: The Ultimate Battlefield of AI Coding

AGI is the ultimate goal for the industry, and AI coding is seen as the shortest path. In early June, Anthropic, the leader in AI coding, announced it had confidentially submitted S-1 draft prospectus to the SEC for IPO, following a $65 billion H-round funding at a post-money valuation of $965 billion, surpassing OpenAI as the highest-valued AI company globally.

Supporting this valuation is strong revenue. Anthropic CEO Dario Amodei disclosed that since its first revenue, the company has grown about 10x annually, with annualized revenue exceeding $44 billion in May 2026. A senior practitioner in China's fintech sector also said, 'I usually use Claude in my daily work, but due to the ban, I sometimes choose Zhipu. AI coding is a must-have efficiency tool; for a team of dozens, monthly spending of hundreds of thousands is normal.'

Notably, AI coding is a typical winner-take-all market where users are extremely sensitive to performance and switching costs are very low. According to UBS research, Moonshot focuses on multi-agent collaboration, MiniMax emphasizes full-stack multimodal capabilities, and DeepSeek specializes in mathematical reasoning. Zhipu, in contrast, is betting heavily on AI coding, publicly stating: 'Since early 2025, we have invested almost all our resources in coding.'

This choice aligns with Anthropic. UBS believes both companies clearly prioritize coding as the core direction for capability frontiers, with relatively lower attention to multimodal technology. According to METR data, Anthropic's Opus 4.6 can achieve about 12 hours of task completion cycle, while Zhipu's GLM-5.1 reaches about 8 hours, ranking first among global open-source models.

Conclusion: Multiple Possibilities on the Path to AGI

Zhipu stated: 'Code is not yet AGI. On the road to AGI, there are more mountains to climb. Our next target is fully autonomous agent systems.' From its strategic intent, the goal is to build, based on long-horizon tasks, a new productivity form comprising self-driving, collaborative, 24/7 operating agent groups—evolving from 'intelligent assistants' to 'digital employees,' constructing an agent society with thousands of different professional 'personalities' and 'skills.'

In the global AI industry's route battle, whether the 'Cathedral' closed-source model or the 'Bazaar' open-source ecosystem will prevail remains unknown. But one thing is certain: when technology blockades become geopolitical bargaining chips, openness, autonomy, and controllability of the open-source model become an unignorable competitive edge. In this race toward AGI, Zhipu's rise may be a vivid preview of historical progress.