Chinese AI models gain favor with U.S. companies as costs soar

Chinese-made AI models are attracting attention among U.S. companies because they can narrow the performance gap with larger U.S. rivals while offering significantly lower usage fees.

Recently released models from Chinese companies such as DeepSeek and Z.ai are seen by many to be more competitive compared to major frontier systems such as Anthropic and OpenAI. Advances in these capabilities come as token prices for cutting-edge models at many U.S. AI labs rise, leaving companies saddled with unexpectedly high costs associated with using the technology.

The share of tokens used by US companies on Chinese AI models via OpenRouter (a platform that allows developers to access various AI models) has been hovering above 30% every week since February 8, with that number rising to 46%. The average for the past 12 months was just 11%, but it dropped to 4.5% in the first half of 2025.

The rise of China’s open source and open weight models comes as the US government is increasingly concerned with regulating the most powerful AI models and considering ways to prevent the rapid introduction of alternatives from abroad.

OpenAI announced at the end of June that it would restrict the deployment of new model sets in response to government requests. Export restrictions on Anthropic’s Mythos and Fable models were also lifted that same month, following a tense standoff between the Trump administration and the company.

“Chinese AI models are particularly attractive to U.S. companies now that AI costs are rising,” Kyle Chan, a fellow at the John L. Thornton China Center at Brookings think tank, told CNBC. “Previously, U.S. companies prioritized AI implementation regardless of model, but now they are focusing more on cost.”

As companies deploy AI models to develop new products and improve internal efficiency, engineers are increasingly experimenting with cheaper open-source and open-weight models, the best of which are made by Chinese companies.

Open source and open weight models allow developers to inspect, use, and possibly modify different parts of an AI model. These are OpenAI, Anthropic, and googlethe code and internals remain proprietary.

In June, AI startup Lindy moved 100% of its traffic from Anthropic’s Claude model to DeepSeek. DeepSeek is a Chinese company that came out with a shocking release in early 2025, announcing a new model in April.

“We did that, and we saw the cost curve come crashing down into the ground,” CEO Flo Crivello told CNBC. He said the decision could save Lindy millions of dollars in the coming months.

DeepSeek saw its share of gateway tokens rise from May to June on Vercel, a platform that allows developers to deploy and run apps and websites.

Z.ai’s GLM 5.2 was widely released in June and will be the earliest to be deployed of any model tracked by Vercel in 2026, Harpreet Arora, head of agent infrastructure at Vercel, told CNBC. “In the first full week after launch, the daily token volume increased by approximately 27 times and the number of customers using the tokens increased by approximately 80 times.”

“Price is doing his job here,” Arora said. “When a task doesn’t require the best model, teams are starting to route to the cheapest model that works well enough, and the recent wave of models coming out of China are winning that deal.”

Justin Summerville, head of data and analytics at OpenRouter, told CNBC that the open source Chinese model could be “60% to 90% cheaper” than the leading Anthropic or OpenAI models.

OpenAI and Anthropic have been contacted for comment.

Cien Solon, CEO and founder of LaunchLemonade, told CNBC that while Claude and ChatGPT still dominate in terms of usage on LaunchLemonade, an AI agent platform for regulated industries, GLM 5.2 is now among the platform’s top five models.

“Z.ai ya (alibabaQwen is becoming a choice for enterprises because it offers an attractive combination of performance and cost for certain workloads. ” Cien said, “Companies with more mature AI strategies are increasingly willing to use Qwen where it makes technical or commercial sense.”

The performance of China’s AI models has also improved.

Although often “a fraction of the cost” of its U.S. competitors, Brookings’ Chan said it operates “close to the top-of-the-line U.S. Frontier model,” and estimates it is currently “six to nine months” behind its top U.S. competitors.

“The new open source model has shown good performance and has proven capable of handling all but the most complex LLM tasks,” Summerville said.

GLM 5.2 came within 1 percentage point of Anthropic’s Opus 4.8 on one notable agent benchmark, at about one-fifth the cost. Some researchers say GLM 5.2 can perform on par with top US laboratories on some cyber benchmarks.

Switching to DeepSeek V4 improved performance for many of Lindy’s key use cases, Crivello said in a post about X.

“Companies are increasingly willing to turn to cheaper AI stacks that they can control and adapt themselves, and given the current state of open-source and open-weight models, that often means leveraging Chinese options,” Yasin Jarnait, head of machine learning at Hugging Face, told CNBC.

“If users want to save money or own their AI stack, there is a real risk that they will have to choose between a high-performance but expensive U.S.-proprietary model, where prices and accessibility can change quickly, or a Chinese model as the only viable alternative.”