Chinese models lead in weekly global AI usage

China's large language models have outpaced their counterparts from the United States in weekly usage for 22 consecutive weeks, firmly holding the top spot globally, according to the latest data from artificial intelligence model aggregation platform OpenRouter.

Chinese AI models processed 62.22 trillion tokens from Sept 21 to Sunday compared with 14.2 trillion generated by US models. A token is the basic unit of text processed by an AI model, and it has become a measure of AI usage and computing consumption.

OpenRouter's data shows the gap was not driven by a single Chinese model, as three of the five most-used models on the platform last week came from Chinese developers.

DeepSeek's V4.1 Flash held the top spot — for a second week in a row — with 19.6 trillion tokens, up 24 percent from the previous week. Z.ai's GLM 5.3 Flash ranked second with 16.3 trillion tokens, up 16 percent. Tencent's Hy4 preview came in fourth with 9.64 trillion tokens.

Hu Yanping, a distinguished professor at the Shanghai University of Finance and Economics, said the rise of Chinese models "reflects a combination of broader model choices, competitive pricing, open-source availability and rapid iteration".

"For companies selecting large language models, the price-performance ratio and whether a model works well with AI agents, tool calling and other enterprise applications are becoming increasingly important alongside raw intelligence. Chinese developers have also been releasing multiple models targeted at different workloads, helping meet a wider range of enterprise demand," Hu said.

Space Bunny Alpha, a model from an anonymous developer, which appeared on OpenRouter on Wednesday, shot to the third spot with nearly 14 trillion tokens in less than a week, ahead of Tencent and OpenAI models.

Space Bunny Alpha offers a context window of 1 million tokens and accepts text, image and video inputs, with adjustable reasoning effort and tool-calling capabilities. While its developer remains anonymous, some foreign independent tests have fueled speculation that it was built by Chinese AI startup MiniMax.

Neither MiniMax nor OpenRouter has officially confirmed the developer's identity. If the model has indeed been made by MiniMax, Chinese models could account for an even larger share of the latest global usage rankings, industry experts said.

Zhong Xinlong, director of the AI research office at the China Center for Information Industry Development's Future Industry Research Center, noted that Chinese model developers' advantages are "extending beyond individual model metrics into full-chain engineering capabilities".

Chinese developers have become faster at moving models from testing phase to their formal release, migrating users to new interfaces and distributing them across multiple platforms, allowing technical improvements to translate rapidly into actual developer usage, Zhong said.

"That engineering speed, coupled with lower inference costs, has become particularly important as AI shifts toward agents that autonomously write code, search databases and call external tools — workloads that can consume far more tokens than a conventional chatbot conversation," he said.

However, usage is not necessarily the same thing as technological leadership or commercial success, Zhong said, adding that the next challenge for Chinese AI companies is turning their cost advantages into reliable services, stronger developer ecosystems and higher-value enterprise applications.