Chinese AI models gain ground in the U.S. as OpenAI and Anthropic costs soar

The CNBC article (July 7, 2026, by Kai Nicol-Schwarz) documents a significant shift in enterprise adoption of artificial intelligence in the United States: American companies are increasingly turning to AI models built in China—such as DeepSeek and Z.ai—as these narrow the performance gap with the frontier systems of OpenAI and Anthropic while proving substantially cheaper to use.

The central figure the piece provides comes from OpenRouter, a platform that lets developers access various AI models: the share of tokens that U.S. companies consume on Chinese models has stayed above 30% every week since February 8, peaking at 46%. To put this in perspective, the average over the previous twelve months was just 11%, and in the first half of 2025 that figure dropped to as low as 4.5%. In other words, in little more than a year the usage share of Chinese models has multiplied several times over.

This shift is occurring against a backdrop of rising per-token prices at many of the most advanced U.S. labs, which has left companies grappling with higher-than-expected AI costs. According to Kyle Chan, a researcher at the John L. Thornton China Center at the Brookings think tank, quoted in the article, "Chinese AI models are particularly attractive to U.S. companies now that AI costs are soaring," and he adds that, whereas companies previously prioritized AI adoption regardless of the model, they are now far more cost-sensitive.

The article frames this surge in Chinese open-source and open-weight models at a moment when the U.S. administration is seeking to regulate its most powerful AI models and assess how to slow the rapid adoption of foreign alternatives. It notes that, in late June, OpenAI agreed to limit the deployment of a new set of models at the government's request, and that export controls on Anthropic's Mythos and Fable models were also lifted that month, following a tense standoff between the Trump administration and the company.

As for concrete adoption cases, the AI startup Lindy shifted 100% of its traffic in June from Anthropic's Claude models to DeepSeek, the Chinese company that burst onto the scene with a surprise launch in early 2025 and released a new model in April. Its CEO, Flo Crivello, told CNBC: "We did it, and you could see the cost curve come down, it literally plummeted." Crivello estimated the decision will save Lindy millions of dollars within a few months, and in a post on X he noted that the switch to DeepSeek V4 even improved performance across many of the company's core use cases.

The report also draws on data from Vercel, a platform for deploying applications and websites: DeepSeek's share of tokens handled through its gateway rose between May and June, and Z.ai's GLM 5.2 model, launched in June to great anticipation, recorded the fastest adoption of any model Vercel tracked in 2026. Harpreet Arora, head of agentic infrastructure at Vercel, explained that in its first full week after launch daily token volume grew about 27-fold and the number of customers using it multiplied roughly 80-fold. Arora sums up the dynamic with a line quoted in the article: "Price is what's doing the work here. When a task doesn't need the best model, teams start routing it to the cheapest model that's good enough, and the recent wave of models coming out of China is winning that bet."

On the cost difference, Justin Summerville of OpenRouter indicates that Chinese open-source models can be between 60% and 90% cheaper than the leading models from Anthropic and OpenAI (both companies were contacted by CNBC for the article, with no response recorded). The piece also quotes Cien Solon, CEO and founder of LaunchLemonade, an AI-agent platform for regulated sectors, who notes that while Claude and ChatGPT still dominate usage on his platform, GLM 5.2 is already among the five most-used models. According to Solon, "Chinese models like those from Z.ai and Qwen (from Alibaba) are becoming options for companies [because] they offer an attractive combination of performance and cost for specific workloads," and he adds that companies with more mature AI strategies are increasingly willing to use them when it makes technical or commercial sense.

In terms of performance, the article notes that Chinese models are approaching the frontier: Chan, of Brookings, estimates they operate "close to the leading U.S. frontier models," lagging the U.S. leaders by "six to nine months," despite costing only "a fraction" of what those cost. Summerville agrees that "the new open-source models are performing well and prove capable for all but the most complex LLM tasks." As a specific benchmark data point, it is mentioned that GLM 5.2 came within less than a percentage point of Anthropic's Opus 4.8 on a well-known agentic benchmark, at roughly one-fifth of the cost, and that some researchers have noted GLM 5.2 can perform on par with the leading U.S. labs on certain cybersecurity benchmarks.

Finally, the article includes a warning from Yacine Jernite, head of machine learning at Hugging Face, about the structural risk this poses: "There is a real risk that users end up trapped having to choose between high-performing but expensive U.S. proprietary models, whose price and accessibility can fluctuate rapidly, or using Chinese models as the only viable alternative when they want to control costs or own their own AI infrastructure." This idea closes the article with a reflection on how price competition between the U.S. and China could, paradoxically, leave Western companies with fewer real options for controlling their AI technology stack, precisely as U.S. regulators try to restrict access to the country's most advanced models.