Chinese AI models, with prices significantly lower than their American counterparts, are rapidly penetrating the US enterprise market. The combination of cost-effectiveness and performance that is increasingly catching up is driving more US developers and companies to shift their workloads to Chinese models.
On the developer platform OpenRouter, the proportion of tokens used by US companies for Chinese AI models has remained above 30% weekly since February 8th, with the peak once climbing to 46%. The average for the previous 12 months was only 11%, and it was even lower at 4.5% in the first half of 2025. This sharp change occurred against the backdrop of leading US AI labs like OpenAI and Anthropic continuously raising prices for their flagship models, accelerating the shift for companies with growing cost sensitivity.
Cost Pressure Drives the Shift
Price is the core driver of this migration wave. According to Justin Summerville from the OpenRouter data and analysis team, using Chinese open-source models is "60% to 90% cheaper" than leading models from Anthropic and OpenAI.
AI startup Lindy is a typical case of this trend. The company shifted all its traffic from Anthropic's Claude model to DeepSeek in June 2026. "You could just see that cost curve just fall through the floor," Lindy CEO Flo Crivello told CNBC. He expects this decision to save the company millions of dollars within months.
Kyle Chan, a researcher at the Brookings Institution's John L. Thornton China Center, noted: "As AI costs soar, the appeal of Chinese AI models to US companies has significantly increased. Previously, US companies didn't scrutinize the origin of models in AI adoption; now their cost awareness is growing stronger."
Performance Rapidly Nears the Frontier
Beyond the price advantage, the performance of Chinese models is also turning heads in the market.
The GLM 5.2 model released by Zhipu in June scored within less than one percentage point of Anthropic's Opus 4.8 on a widely watched agent benchmark, while costing only about one-fifth as much. Some researchers noted that GLM 5.2's performance on certain cybersecurity benchmarks is already comparable to that of top US labs.
Harpreet Arora, head of agent infrastructure at Vercel, stated that GLM 5.2 set the fastest model adoption rate Vercel tracked in 2026—"in the first full week after launch, average daily token volume grew about 27x, and the number of customers using it grew about 80x."
According to Vercel data, DeepSeek's share of gateway tokens also rose steadily between May and June.
Summerville said the new batch of open-source models "are performing very well and have proven to be capable for practical use except for the most complex LLM tasks." Chan estimates that China's top models currently lag behind the US frontier products by "about six to nine months," but are already sufficiently competitive for many mainstream scenarios.
Enterprise Adoption Widens
The penetration of Chinese models is extending from startups to a broader range of enterprises.
On LaunchLemonade, an AI agent platform for regulated industries, Claude and ChatGPT still top the usage charts, but GLM 5.2 has entered the platform's top five. LaunchLemonade founder and CEO Cien Solon said: "Chinese models like Zhipu and Alibaba's Qwen are becoming viable options for enterprises, offering an ideal combination of performance and cost for specific workloads. Companies with more mature AI strategies are increasingly willing to use them in scenarios where the technology or business logic justifies it."
Yacine Jernite, machine learning lead at Hugging Face, pointed to a deeper structural trend: "We see companies increasingly wanting to move to AI stacks that are cheaper, that they can control and customize themselves, and given the current state of open-source and open-weight models, that often means turning to Chinese models."