US companies’ use of Chinese artificial intelligence services surged in June, when Anthropic suspended some models over government restrictions, as open-source models with prices as low as one-twentieth of Anthropic’s Claude Mythos become more entrenched in the American corporate world.
In response to soaring fees, US cryptocurrency exchange Coinbase shifted much of its AI use to cheaper Chinese models as GLM and Kimi. This and other steps have “cut our AI spend nearly in half,” CEO Brian Armstrong wrote in a social media post on June 27.
GLM, developed by Z.ai, and Kimi, developed by Moonshot AI, both excel at software coding and time-consuming tasks. Their cost per token is about one-twentieth that of Anthropic’s latest models.
GLM-5.2, released in mid-June, is drawing notice for its performance. In testing with roughly 9,000 problems to measure the practical capabilities of AI, it ranked fifth out of around 500 models, surpassing those from Google and others.
Many US AI experts have likened it to China’s DeepSeek, whose R1 model appeared in January 2025 and stunned the world with its capabilities and low costs.
Since the 2022 public release of OpenAI’s ChatGPT, American companies have led the way in AI performance, with Chinese companies following closely behind. Most US AI models are closed, meaning that their technology is kept secret, while many Chinese companies build inexpensive and easily replicable open-source models.
As AI evolves, soaring usage fees have emerged as a hurdle. More complex tasks lead to the use of more tokens, pushing up costs. Some major US companies, including Airbnb and Uber Technologies, have acknowledged internal use of Chinese AI.
US President Donald Trump’s administration, previously reluctant to regulate AI, imposed export controls on Anthropic’s Claude Mythos 5 and Claude Fable 5 in June, resulting in the company temporarily suspending both models.
The administration’s shift to a more controlling stance dramatically increased the risk of sudden action making services unavailable, leading more companies to adopt Chinese models.
Data from US startup OpenRouter—which offers a service used by more 8 million users, primarily engineers, that allows them to switch freely between more than 400 types of AI—was used to compare the number of tokens consumed by US and Chinese AI.
Chinese models, which had less than half the number of tokens used as US models at the start of the year, surpassed US models multiple times after February. Their use exploded in June, reaching 25 trillion tokens in the final week of the month, more than double the number at the end of May and 78% more than US models.
DeepSeek held the top spot in market share by company in June, at 19%. Such other Chinese companies as Z.ai and Xiaomi also saw significant growth, while Google’s and OpenAI’s shares shrank.
“With cost effective token usage dominating the summer headlines, we’d expect these cheaper-but-still-powerful [Chinese] open source models to continue to attract more tokens over time,” OpenRouter said in a report.
The US is increasingly wary of the rapid advances in Chinese AI. Anthropic CEO Dario Amodei has posited that AI will be used to strengthen authoritarian states like China, including through surveillance of citizens, leading to a decline in democracy.
A potentially greater threat is Chinese AI companies engaging in distillation, using the output of high-performance AI models to develop other AI, reducing costs while maintaining performance. Many US companies prohibit unauthorized distillation.
In mid-June, Anthropic wrote to US senators to demand action against Alibaba, accusing the Chinese technology group of “the largest known distillation attack” on the company to date.
If distillation becomes persistent, even if US companies continue to invest massive funds to develop cutting-edge AI, Chinese companies will be able to quickly catch up.
Using open-source AI models comes with its own risks, as there is no guarantee that Chinese companies will continue to provide their models to external parties indefinitely. There is also a chance that high-performance Chinese AI comparable to Claude Mythos could be developed.
“Open-source models are difficult to regulate with suspensions or prescreening measures because they can be replicated,” said Boston Consulting Group’s Shinichi Takayanagi, an expert on AI. “If they are available at a low price, it will put pressure on US AI companies to lower prices too, resulting in insufficient investments and potentially leading to neglect of AI safety measures.”
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.