Chinese Models Take More Than 60% of OpenRouter Tokens, Driven by Lower Prices

Chinese-developed models have not been shown to lead global AI traffic. But on OpenRouter, a platform built to let developers compare and route work across models, they accounted for more than 60% of tokens by June, while U.S. models fell below 40%. The shift puts the economics of high-volume AI work in focus.

OpenRouter serves more than 8 million developers through a common API. Its data offers a useful look at a market where switching models is relatively easy and price-performance is exposed, but it does not establish worldwide AI-usage shares.

The lead held through July

A China Europe International Business School analysis found Chinese models first moved ahead of U.S. models in weekly OpenRouter token use during the first two weeks of March 2026. By June, OpenRouter data put Chinese models at about 18 trillion weekly tokens, versus roughly 5.5 trillion for American models. In July, Chinese-developed models occupied all five leading OpenRouter positions by token volume, led by Xiaomi’s MiMo V2.5.

18 trillionChinese models

OpenRouter data put Chinese models at about 18 trillion weekly tokens in June.

5.5 trillionAmerican models

The same June comparison put American models at approximately 5.5 trillion weekly tokens.

Cheap output favors token-heavy work

DeepSeek, Alibaba, Moonshot AI, MiniMax, and Zhipu AI have released increasingly capable models at lower prices. One comparison cited DeepSeek V3.2 at $0.42 per million output tokens, against $75 for Anthropic’s Claude Opus. OpenRouter’s Justin Summerville estimated Chinese open-source models run 60% to 90% cheaper than leading Anthropic and OpenAI offerings.

The difference grows more important as automated workloads consume more output. Research cited by CEIBS found programming rose from 11% of model use in early 2024 to more than half, while agent-based workflows generated more than half of output tokens. Airbnb also used Alibaba’s Qwen for customer service, describing it as fast and cheap in material referenced by a congressional letter.

Distribution extends the cost advantage

• Open-weight models can be downloaded, fine-tuned, and self-hosted or served through third-party infrastructure.
• Hugging Face CEO Clément Delangue said Chinese-developed models account for 41% of downloads on the repository. Fireworks, Together, and DeepInfra also host Chinese models.

Deployment does not erase governance questions

DeepSeek’s privacy policy says personal data from its services is collected, processed, and stored in the People’s Republic of China. It says the company may share personal data with law enforcement or public authorities when necessary to comply with law or legal process.

The House Homeland Security Committee and House Select Committee on China opened an investigation into PRC-developed models, naming DeepSeek, Alibaba, Moonshot AI, and MiniMax and raising provenance, censorship, cybersecurity, and supply-chain concerns. Senator Tom Cotton separately urged a government-wide ban on contractors’ use of Chinese AI models in a July 29 letter.