White House AI Clampdown: A Boon for Chinese Model Makers

The recent regulatory scrutiny faced by leading U.S. artificial intelligence developer Anthropic appears to be creating an unintended advantage for China, a key rival in the global AI race. After a two-week suspension stemming from an export control directive, Anthropic received White House approval on Friday to release its advanced Mythos 5 model to select companies and federal agencies. However, its Fable 5 model remains unavailable. Concurrently, OpenAI announced on Friday it would also restrict the deployment of its GPT 5.6 models following a government request.

These developments occur as Anthropic and OpenAI, along with tech giants like Google, engage in a fierce competition to develop the most sophisticated AI technologies. The U.S. government has historically favored a less restrictive regulatory approach to foster rapid AI development, a stance many industry executives and administration officials believed was crucial to preventing China from closing the technological gap.

Yet, as U.S. AI leaders navigate national security concerns and government directives, Chinese companies are accelerating the release of AI models that demonstrate capabilities competitive with top U.S. research labs. Notably, Zhipu’s GLM 5.2, launched earlier this month, has been benchmarked by researchers as performing comparably to leading U.S. AI models on certain cybersecurity tasks, even matching the performance of Anthropic’s Mythos model in some areas.

“Many sharp minds in the AI community are observing that GLM-5.2 represents the first Chinese AI model to consistently match, and often surpass, leading public AI models from American labs, without apparent compromises,” noted venture capitalist Marc Andreessen in a recent social media post. He added, “This timing is particularly striking given current geopolitical developments.”

Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, described the current situation as a “significant wake-up call.” In a report to clients, Jefferies strategist Christopher Wood, citing industry sources, highlighted that GLM 5.2 “is nearly on par with Anthropic as a competitor in the corporate market and operates at just one-quarter of the cost per token.”

Even figures who have been critical of Anthropic’s AI safety protocols, such as former Trump administration advisor David Sacks, have pointed to China’s advancements. Sacks shared a post on social media above a headline from The Wall Street Journal stating that China had achieved parity with Anthropic in cybersecurity capabilities.

“A year ago, President Trump articulated a clear strategy for America to win the global AI race by prioritizing innovation, infrastructure, energy, and exports,” Sacks commented. “His foresight was accurate; any deviation from this strategy carries substantial risks.”

Representatives from Anthropic, OpenAI, and the White House did not immediately respond to requests for comment.

This emergence of Chinese AI capabilities coincides with a significant shift in the corporate AI landscape. Many U.S. companies are moving away from an era of unrestrained AI spending, often termed “tokenmaxxing,” towards a more focused approach emphasizing efficiency and return on investment. This trend also inadvertently benefits Chinese AI developers.

Earlier this month, Flo Crivello, CEO of AI startup Lindy, reported a strategic pivot, migrating 100% of her company’s AI traffic from Anthropic’s Claude models to DeepSeek, a Chinese firm offering more cost-effective, open-weight alternatives.

“The transition was immediate, and we observed a dramatic reduction in our cost curve,” Crivello stated in a recent interview. “It was a significant optimization for us.”

Chinese AI developers are gaining traction with U.S. users due to the accessibility of open-weight models. These models can be downloaded and deployed on a company’s own infrastructure, circumventing reliance on third-party cloud services and their associated costs.

“With open-weight models, the landscape is akin to the Wild West,” commented Travis Lanham, co-founder of AI security startup Armadin, which is actively evaluating GLM 5.2 and Kimi K2.7 from Moonshoot AI, another Chinese developer.

Lanham noted that these models are exhibiting enhanced capabilities in cybersecurity applications, including the analysis of reconnaissance data and the generation of exploit code for customer-facing vulnerabilities.

The ongoing question for U.S. policymakers revolves around how to balance national security with the rapid pace of AI innovation, particularly given the disparate approaches to handling sensitive technology between the U.S. and China. For years, the U.S. has implemented stringent export controls on advanced AI chips from companies like Nvidia and Advanced Micro Devices to limit China’s access to cutting-edge hardware. Furthermore, national security concerns led to the U.S. barring American companies from using equipment from Huawei.

While the U.S. did authorize the export of Nvidia’s H200 chip—a model utilized by domestic AI firms—to the China region last year, Nvidia reported earlier this year that it had not yet generated any revenue from these sales and remained uncertain about China’s willingness to import its products.

Regarding the capabilities of GLM 5.2, Elon Musk, founder of Tesla and SpaceX, projected on social media that the model would likely achieve parity with Anthropic’s Fable model by the first quarter of next year. This was in response to a user’s query about the timeline for such a milestone. Zhipu founder Jie Tang replied to Musk, stating, “It won’t take that long.”

The appeal of Chinese AI models extends beyond niche applications, with major companies also adopting them. Shopify and Airbnb, for instance, have publicly recognized the benefits of Alibaba’s Qwen 3 in scaling their AI functionalities. Brian Armstrong, CEO of Coinbase, shared on social media that his company is leveraging open-weight models like GLM 5.2 and Kimi 2.7, which has enabled them to reduce their AI expenditure by nearly half despite an increase in token usage.

Cybersecurity remains a paramount concern for many industry experts. Some open-weight models are already capable of automating multiple stages of a cyberattack, and Hed Kovetz, CEO of industry startup Silverfort, expressed apprehension that these models could orchestrate an entire attack operation within months.

“If the U.S. government does not empower the industry to proactively prepare, then by the time Chinese models reach a comparable level of sophistication, preparedness will be severely lacking,” Kovetz warned.