AI Usage Fees in China Just One-Hundredth of U.S.; "U.S. Companies Also Favor DeepSeek"

Share of Chinese AI Usage on OpenRouter Reaches 46%

It has been observed that there is a growing tendency among U.S. companies to prefer Chinese artificial intelligence (AI) models. This surge in demand is largely attributed to the overwhelming price gap, as the usage fees for Chinese AI models are about 1/100th of those for their American counterparts.

DeepSeek logo. Photo by Reuters and Yonhap News Agency

According to reports from foreign media such as Japan’s Nihon Keizai Shimbun (Nikkei), on OpenRouter, a service that U.S. companies use to integrate AI into business software, the share of Chinese AI usage (measured by tokens) has recently reached 46%. In the first half of last year, the share was only in the 4% range, but it surpassed 30% in February and has continued to grow. In AI models, tokens are the basic units for processing text and are used as a standard for both pricing and usage volume.

As companies increasingly apply AI across all internal services, there has been a surge in demand for the more affordable Chinese AI models. For every one million tokens, the usage fee for Anthropic’s ‘Claude Opus 4.7’ is $25, whereas China’s ‘DeepSeek V4 Flash’ costs only $0.18. This results in a price difference of about 139 times. One million tokens correspond to the amount of input an AI language model can process at once, which is equivalent to roughly 700,000 to 800,000 English words.

Unlike most U.S. AI companies, which primarily use closed-source models, open-source models dominate in China. With closed-source models, costs soar as token usage increases. Z.ai (Zhifu AI), a Chinese AI startup known for the ‘Mini DeepSeek Moment,’ released its ‘GLM-5.2’ model last month, which has already surpassed some of Anthropic’s models in OpenRouter usage rankings. For this reason, U.S. AI assistant startup Lindy switched its AI model from Anthropic to DeepSeek last month.

Recently, a sudden export restriction on Anthropic models by the Donald Trump administration has accelerated this shift. Last month, the U.S. Department of Commerce restricted overseas access to ‘Fable 5’ and ‘Mythos 5’ for security reasons, only to lift the ban three weeks later. The global interest in Chinese models has also increased as the release of OpenAI’s latest model, ‘GPT-5.6,’ has been delayed due to security concerns.

Kyle Chan, a researcher at the Brookings Institution, explained, “Due to export controls, there are growing doubts about whether U.S.-made AI models can be used reliably in the future. In contrast, many Chinese AI models are open-source and can be downloaded to a company’s own servers, making them more stable.” However, Nikkei pointed out that geopolitical risks remain, as the Chinese government is also considering export restrictions on advanced AI models from Alibaba Group and ByteDance.

This content was produced with the assistance of AI translation services.