China’s MiniMax, Moonshot lead AI token usage ranking

Chinese open-source AI models from MiniMax and Moonshot have led token usage on the OpenRouter platform, ending a year of US dominance, according to its latest ranking.

MiniMax’s M2.5 from Shanghai became the most used model so far this month on OpenRouter with 4.6 trillion tokens.

Moonshot’s Kimi K2.5 from Beijing ranked second with 4.0 trillion tokens so far this month on the platform.

Other top models included Google DeepMind’s Gemini 3 Flash Preview, DeepSeek’s V3.2, and Anthropic’s Claude Sonnet 4.5.

The surge in Chinese model usage coincided with intensified development ahead of the Chinese New Year, involving firms like Alibaba Cloud, Moonshot, MiniMax, and Zhipu AI.

A February report by Frost & Sullivan noted that Chinese enterprises’ average daily token usage reached 37 trillion in the second half of 2025, up 263% from 10.2 trillion in the first half.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Aggressive pricing and automated work drive China’s AI surge

Chinese open-source AI models are picking up users by aiming at heavy-use work such as automated coding and “agent” tasks (AI systems that can plan and carry out multi-step work). That shift raises token use as activity moves past simple Q&A 1.
Shanghai-based MiniMax built M2.5 for agent use cases, and its token volume on OpenRouter rose 197% in one week 1.
MiniMax M2.5 also competes on price. Its input cost is more than 16 times lower than Anthropic’s Claude Opus 4.6 input pricing 1.
Anthropic alleges Chinese AI labs including DeepSeek, Moonshot AI, and MiniMax used about 24,000 fraudulent accounts to create over 16 million exchanges with Claude. Anthropic describes the effort as a coordinated “distillation” campaign meant to extract high-value model outputs 2.

The fight for AI dominance moves from chips to API calls

Chinese model momentum opens a new phase in the tech rivalry, where model capability matters as much as the hardware running it.
Claims of industrial-scale “distillation” describe Chinese labs allegedly pulling capabilities from US models through millions of API (application programming interface) queries. That undercuts chip export controls as a stand-alone limit 2.
This trend could split the global market, with a cheaper stack that may ship without the safety guardrails common in Western models 3.
US companies and policymakers may start treating model access as a national security concern, which could lead to limits on advanced AI access similar to controls on sensitive hardware 3.