AI models developed by Chinese companies have, for the first time, surpassed their U.S. counterparts in global token usage.
According to the late-February ranking by OpenRouter, a platform that simplifies interactions with AI models, Chinese developers now occupy several top positions among the world’s most widely used large language models (LLMs). This trend reflects the growing adoption of Chinese AI technologies by developers worldwide.
Specifically, the M2.5 model developed by Chinese startup MiniMax currently leads the global ranking, processing approximately 1.7 trillion tokens per week. The American model Gemini 3 Flash Preview from Google ranks second with around 997 billion tokens, followed by DeepSeek V3.2 with about 798 billion tokens. Other Chinese models, including Kimi K2.5 from Moonshot AI and GLM-5 from Zhipu AI, also appear among the world’s most widely used models, each exceeding 600 billion tokens in weekly usage by developers.
During the week of February 9–15, Chinese AI models collectively recorded 4.12 trillion tokens, surpassing U.S. models for the first time, which processed 2.94 trillion tokens during the same period. In the following week (February 16–22), usage of Chinese models further increased to 5.16 trillion tokens, while U.S. models declined to 2.7 trillion tokens.
Analysts attribute this surge to two primary factors: the increased use of AI during the Lunar New Year period and the broader deployment of AI agents, which significantly increase token consumption per task. OpenRouter confirmed that demand for generating long-form content has grown noticeably in recent weeks, with MiniMax M2.5 leading in workloads consuming 100,000 to one million tokens, a typical range for agent-based AI workflows.