At the recent AI Impact Summit in New Delhi, Dario Amodei, the co-founder and CEO of Anthropic, addressed crucial developments in the artificial intelligence sector. Following a two-week halt prompted by export control regulations from the Trump administration, the White House has permitted Anthropic to launch its advanced Mythos 5 AI model to select enterprises and federal agencies, although its Fable 5 model remains unavailable. Concurrently, OpenAI announced restrictions on the distribution of its GPT 5.6 models in response to a government directive.
As the U.S. tech landscape evolves, Anthropic and OpenAI find themselves in a competitive race with major players like Google to innovate the latest AI technologies. The Trump administration has championed an acceleration of AI advancement by minimizing regulatory barriers, with many tech leaders expressing concern that slowing down could inadvertently benefit China, which has been rapidly advancing in artificial intelligence capabilities.
While Anthropic complies with U.S. government security measures, Chinese firms are making strides with models that rival U.S. outputs. A recent assessment highlighted that Zhipu's GLM 5.2, launched earlier this month, can perform on par with top U.S. laboratories in specific cyber benchmarks, even matching the functionality of Mythos. Venture capitalist Marc Andreessen noted on social media that GLM 5.2 appears to be the first Chinese AI model to compete head-to-head—and often surpass—American models without sacrificing performance.
Experts like Sam Bresnick from Georgetown’s Center for Security and Emerging Technologies view these developments as significant wake-up calls. Jefferies strategist Christopher Wood, referencing industry insights, remarked that GLM 5.2 is nearly on par with Anthropic in the corporate sector while costing a fraction of the expense. Even David Sacks, a former advisor to Trump on crypto and AI, cautioned on social media about China's notable advancements in cybersecurity, echoing Trump’s earlier call for a proactive approach to maintain the U.S. position in the AI race.
Amid this backdrop, American enterprises are shifting from unrestricted AI spending, often termed "tokenmaxxing," to a sharper focus on efficiency and return on investment—a trend that may inadvertently favor Chinese developers. For instance, CEO Flo Crivello of AI startup Lindy recently transitioned away from Anthropic's Claude models to DeepSeek, citing significant cost reductions.
The ease with which Chinese AI solutions can reach U.S. consumers is partly due to companies being able to download open-weight models and operate them independently, thus bypassing third-party cloud dependencies. This has led to increased experimentation with models like GLM 5.2 among startups focused on AI security. However, the future of these developments is uncertain as policymakers consider how to handle the sensitive hardware exchanges between the two largest economies.
Historically, the U.S. has implemented export controls to prevent advanced AI technologies from reaching China. This includes restrictions on AI chips produced by Nvidia and AMD, as well as a ban on American firms utilizing Huawei technology due to national security concerns. In a notable turn, the U.S. government authorized the export of Nvidia's H200 chips to China, but Nvidia has indicated that it has yet to see any revenue generated from these exports.
In a discussion regarding the capabilities of GLM 5.2, Elon Musk speculated that this AI model could soon match the levels of Fable. Zhipu's founder Jie Tang responded optimistically, suggesting that this could happen sooner than expected. Notably, major corporations—including Shopify and Airbnb—are also embracing Chinese solutions, like Alibaba's Qwen 3, to enhance their AI functionalities. Coinbase's CEO revealed that his firm is utilizing open-weight models to decrease AI expenses significantly, even as their token usage increases.
Concerns regarding cybersecurity dominate discussions among industry experts, with several open-weight models potentially capable of automating various stages of cyberattacks. Hed Kovetz, CEO of Silverfort, voiced worries that these technologies could soon execute entire operations. He stressed the urgency for the U.S. government to leverage the current moment to prepare the industry; failing to do so may leave them unprepared when Chinese models achieve similar capabilities. Contributions to this article came from CNBC's Deirdre Bosa and Ashley Capoot.
