For developers outside China, Chinese AI models had become, over the past several months, an option there was simply no reason not to use. Their performance approached that of frontier US models, yet their prices were an order of magnitude cheaper. Now, however, Reuters has learned through independent reporting that China itself—the source of that cheapness—is moving to restrict overseas access to its own top-tier models, including DeepSeek and Qwen. Just weeks after the US shut off foreign access to Anthropic's flagship models, China, the supplier in this equation, has begun moving in the same direction.
Ministry of Commerce-Led Closed-Door Talks Target DeepSeek, Qwen, and GLM-5.2
According to a Reuters report dated July 7, 2026, China's Ministry of Commerce has spent roughly the past month holding closed-door discussions with Alibaba, ByteDance, and the startup Z.ai (Zhipu AI). The ministry, which oversees foreign trade and export controls, reportedly had officials from the National Development and Reform Commission (NDRC) sitting in on these talks as well. The information comes from three anonymous sources, and Alibaba, ByteDance, and Z.ai have all declined to comment.
Reuters reports that the discussions cover not only already-released models—DeepSeek, Qwen, and GLM-5.2—but also overseas access to as-yet-unannounced frontier models, whether closed-source or open-source. The scope and timing of any measures remain undecided. According to Reuters, it is not yet clear whether the restrictions would extend to models already deployed overseas, such as DeepSeek, Qwen, and GLM-5.2, or whether they would apply only to future models yet to be announced. If models with existing overseas user bases were included, the impact would extend to the entire base of users of Chinese models on OpenRouter. The fact that Alibaba, ByteDance, and Z.ai—the very companies at stake—have all remained silent also suggests the matter is still confined to internal deliberation.
DeepSeek has already been banned from use by government agencies in the US federal government, Australia, Taiwan, Italy, and elsewhere. This latest development means that China itself—the sender's side of that equation—has begun considering an option in which it withholds its own flagship models from overseas markets. The caution that had previously taken shape on the receiving end is now beginning to take shape on the sending end as well.
Why China Can Legally Restrict Its Own AI: The Three-Tier Regulatory Concept
Reuters reports that in May 2026, a summary of discussions from a gathering of Chinese legal experts was published in a journal affiliated with the Supreme People's Court, outlining a proposal to manage AI models across three tiers. Under this framework, basic open-source models would be subject to a registration system, more advanced technologies would undergo safety review, and the most sensitive frontier models would be kept confidential or restricted to domestic use only. While the original document itself could not be directly confirmed, this proposal aligns in direction with the current talks led by the Ministry of Commerce and NDRC, and could serve as the legal foundation for China to halt overseas deployment of its own top models.
The Ministry of Commerce and NDRC talks are also said to have touched on proposals to treat leaks or theft of AI technology as violations under national security law, as well as proposals to restrict who can invest in domestic AI startups. In early June 2026, new regulations were also introduced targeting overseas transactions involving Chinese investors, technology, and data—a parallel move to tighten outflows of capital and technology alongside AI itself. Backing up this enforcement capacity is the action NDRC took against Meta in April 2026, forcing the unwinding, within a set deadline, of its roughly $2 billion (approximately ¥322 billion at the July 2026 exchange rate of $1 = ¥161) acquisition of the Chinese startup Manus.
This containment effort began even earlier. Bloomberg reported that as of May 26, 2026, AI engineers at private companies including Alibaba became subject to a government permit requirement for overseas travel. In other words, China restricted the movement of talent in May, regulated cross-border transactions of capital and data in early June, and by June had begun considering restrictions on overseas provision of the models themselves. Within a span of less than three months, China has sealed off outflow channels for AI-related resources one by one, in the order of people, capital, and technology.
The Math Behind How Regulation Would Erase a $544-per-Task Price Advantage
Z.ai's GLM-5.2 has managed to approach frontier-model performance while cutting prices sharply. According to industry estimates cited by CNBC, in one comparison of the same task, Anthropic's Claude cost $4,811 and OpenAI's ChatGPT cost $3,357, while Zhipu AI's GLM completed the task for just $544. That's a difference of $4,267 versus Claude—putting the cost roughly one-ninth as high.
According to data compiled by the-decoder, the weekly traffic share of Chinese models via OpenRouter has surged from an average of 11% the previous year to consistently exceed 30% in recent weeks, hitting a record 46% in one particularly notable week. If that 30% share of calls were to shift entirely to Claude, costs would balloon roughly 8.8-fold, from $544 to $4,811 per task. For a development team running 1,000 tasks a month on GLM and paying $544,000, a shift to Claude would push that bill to $4,811,000—a difference of roughly $4.27 million. What this calculation illustrates is the scale of cost increase development teams could face if such restrictions take effect, with the impact growing larger the more heavily a company depends on GLM.
Just Weeks After Restricting Anthropic, the US and China Are Simultaneously Locking Down Their Own AI
In June 2026, the US administration moved to ban foreign-national users from accessing Anthropic's flagship models, Fable and Mythos. Because there was no immediate way to verify user nationality, the restriction briefly resulted in a worldwide service suspension. Fable was later restored after enhanced safety measures were implemented, but Mythos remains limited to a select group of "trusted" US organizations.
Mythos carries strong characteristics as an AI tool for cybersecurity, and there is reportedly growing concern within China that it could be repurposed for attacks against China. Zhou Hongyi, founder of the security firm 360, has reportedly publicly called for the need for a "Chinese version of Mythos" as a countermeasure. That China began considering similar measures just weeks after the US locked down its own AI seems less like coincidence and more like the natural result of mutual wariness reaching a critical point at roughly the same time.
That said, the two moves point in opposite directions. The US measure was a "usage restriction"—stopping foreign-national users from using America's own frontier models. China's proposed measure, by contrast, is an "export control"—stopping its own companies from providing their models overseas, with the restriction aimed at the supply side rather than the demand side. Precisely because what each side is trying to protect points in opposite directions, both countries have arrived, at nearly the same moment, at the same conclusion: keeping their top models within their own borders.
Mistral—one of Europe's few homegrown frontier-model developers—along with major US players like OpenAI and Anthropic, whose pricing pressure may ease, stand to benefit relatively if cheap Chinese alternatives exit the market. Within China, security firms like 360, which stand to capture demand for a Mythos alternative, also gain a tailwind. On the other hand, overseas startups that had fully migrated to Chinese models like DeepSeek to cut costs, Alibaba, ByteDance, and Z.ai themselves—who stand to lose overseas expansion opportunities—and Europe, which lacks its own AI infrastructure, all find themselves squeezed between these opposing forces, with their options narrowing. This symmetrical lockdown is beginning to divide winners from losers.
According to a report from the European Investment Bank (EIB), lead investors in more than four out of five major EU funding rounds are foreign, in stark contrast to the US, where local, San Francisco-based investors lead only 14% of rounds. For Europe, which lacks deep pools of its own capital and its own models, the disappearance of cheap Chinese AI as an option would leave little room to find a substitute. Development teams anywhere—including in Japan—that have built their API integrations around GLM's cost advantage cannot consider themselves immune to the kind of $4.27 million cost increase such regulation could trigger.
Before the Details Are Even Settled, Some Are Already Shifting from Choosers to the Excluded
Neither Alibaba, ByteDance, and Z.ai, nor the Ministry of Commerce and NDRC have commented on the matter, meaning the only confirmed fact so far is what Reuters obtained from three anonymous sources: that discussions are taking place. The next indicators to watch will be whether the Ministry of Commerce releases formal regulatory documents, and whether the three companies involved address the matter in earnings calls or in responses to regulators.
In June, the US locked foreign users out of its own frontier models; just weeks later, China began moving in the same direction. As each side, wary of the other, moves to wall off its own AI assets, the ones left paying the price are not the parties to the conflict themselves, but the European and third-country developers who had positioned themselves to choose cheap, capable AI. China's sequential sealing-off of outflow channels—people, then capital, then technology—signals that the struggle for AI dominance has entered a phase where it's treated as part of broader trade friction and technological rivalry.
What happens next will hinge on how much of the three-tier regulatory framework China actually codifies into law, and whether it extends the scope to existing models. Whether DeepSeek and Qwen actually disappear from APIs, or whether this remains merely a matter "under consideration," will come down to a single decision by the Ministry of Commerce this summer.