Zhipu co-founder Tang Jie is arguing that powerful AI should stay open just as U.S. controls have made access to Anthropic's strongest models a political decision. You don't need to buy every part of his case to see why developers are listening.
Tang Jie didn't mince words. When Washington pushed Anthropic to cut off foreign access to Claude Fable 5 and Mythos 5 on June 12, he called the move "deeply regrettable" in comments reported by CNBC. That was the right phrase. A model that can disappear from your workflow overnight isn't just a product risk. It's a sovereignty risk, and anyone building on someone else's closed system has to price that in now.
This Is An Access Fight Now #
Anthropic's shutdown was not rumor or diplomatic theater. The Verge reported that Anthropic pulled access to Fable 5 and Mythos 5 after a U.S. government directive citing national security concerns. AP later reported that the broad restriction was lifted at the end of June, with Fable 5 back in wider use while Mythos 5 remained limited to selected U.S.-approved organizations. That detail matters. The door didn't stay shut for everyone, but Washington showed it can decide who gets the key.
Zhipu, which now brands itself internationally as Z.ai, answered with the opposite move. Around the same period, it released GLM-5.2, a 744 billion parameter open-weight model, under an MIT licence. Anyone can download it, run it, modify it, and sell products built on top of it. That's the whole point. If you're a developer in Singapore, Berlin, São Paulo, or Lagos, the appeal isn't ideological. It's practical. You can build without waiting for a compliance email from San Francisco or Washington.
The performance claims are why this isn't just open-source theater. Barron's reported this week that GLM-5.2 ranks fifth globally in benchmark comparisons and ahead of Google's Gemini models, while its hosted version in China costs about 15% of OpenAI's model pricing. TechRadar separately noted that GLM-5.2 topped a Design Arena HTML web design leaderboard and priced tokens at $1.40 for input and $4.40 for output per million tokens, far below Claude Fable 5's $10 and $50. Those numbers will get attention. They already have.
Here's the thing: openness does not make safety simple. Tang's argument is that public participation and open testing, backed by transparency, are safer than closed access controlled by a few companies and governments. Anthropic's argument is that frontier systems can be abused at scale, and that open access gives bad actors a shorter path. Both sides have evidence under them. Only one side is pretending the market will wait for the policy debate to settle.
Anthropic has real reasons to be angry. The Financial Times reported that, in a June 10 letter to the Senate Banking Committee, Anthropic accused Alibaba-affiliated operators of using about 25,000 fraudulent accounts to generate more than 28.8 million Claude exchanges between April 22 and June 5. Anthropic called it the largest known attempt to extract Claude's capabilities. Alibaba has not publicly accepted that account. Still, if you're Anthropic, you don't look at those figures and conclude that borderless access is harmless.
But if you're Zhipu, you look at the same fight and see an opening. Closed American models are becoming more expensive, more regulated, and more politically exposed. GLM-5.2 is cheap, downloadable, and difficult to switch off by order. That combination is powerful. It doesn't have to beat every U.S. model on every benchmark to change buying behavior. It just has to be good enough, cheap enough, and available when the closed model isn't.
The Stock Move Says It All #
The market has made its own judgment. The Wall Street Journal reported that Zhipu's shares rose after the company announced a plan to raise about $4 billion through a discounted share sale, with proceeds aimed at research and development, cloud services, operations, and possible acquisitions. Barron's put the stock gain since its January Hong Kong listing at more than 1,000%. That's not normal. It's a bet that Chinese open-weight models are moving from developer curiosity to commercial threat.
Zhipu listed in Hong Kong on January 8 under the formal name Knowledge Atlas Technology. Its backers and customers are not operating in a neutral environment. The company has been on the U.S. Entity List, and Washington's chip controls still hang over every Chinese AI lab trying to train and serve frontier models. Yet GLM-5.2's pitch is partly that it can run on Chinese hardware, including Huawei's Ascend ecosystem. That doesn't solve every compute constraint. It does make the company less dependent on Nvidia shipments moving cleanly through U.S. policy.
Frankly, Washington's policy does not look settled. The U.S. is trying to stop Chinese labs from accessing the most capable American models and chips, while developers around the world are discovering that Chinese open-weight systems may be cheaper and easier to control on their own infrastructure. You can ban an API endpoint. You can't recall a model that has already been downloaded and mirrored.
Zhipu's bet is clear. Give developers the model. Win the distribution. Figure out the economics later, through usage, cloud services and enterprise demand. Anthropic's bet is different: keep the strongest systems guarded enough that regulators and national security customers stay comfortable. Neither path is clean. But if your startup has to ship this quarter, the open model that works today has a brutal advantage over the closed model that might need political permission tomorrow.
Also read: How to Evaluate AI Agents Before You Ship Them to Real Users • Empery Digital Dumps Nearly Half Its Bitcoin to Chase an AI Data Center Bet • What Is an AI Wrapper Startup and Why VCs Are Suddenly Skeptical