February 8 marked a significant shift in the American landscape for AI developers. Recent statistics show that models developed in China now account for over 30% of token usage among US users on OpenRouter, a well-known API aggregation platform that facilitates the switching of various large language models. This percentage has consistently remained above 30%, hitting a peak of 46%. For comparison, these models previously averaged only around 11% of the US token usage in the past year.
The rise of Chinese AI models did not just begin in February. According to a collaborative analysis by OpenRouter and Andreessen Horowitz, these models had seen a dramatic increase from merely 1.2% of platform token usage in late 2024 to almost 30% throughout various points in 2025, based on an evaluation of 100 trillion tokens. By late February, a new high was reached, with Chinese models capturing a striking 61% of total token volume among the top 10 models on OpenRouter.
Leading this surge was MiniMax M2.5, which achieved an impressive 2.45 trillion tokens used during that period. Other notable performers included Moonshot AI’s Kimi K2.5 and Zhipu AI’s GLM-5.
Understanding the Cost Advantage of Chinese AI Models The underlying factor driving this trend is primarily economic. Chinese-built AI models significantly undercut the prices of comparable offerings from established companies such as OpenAI and Anthropic, offering a reduction of between 60% and 90% on a per-token basis.
These models are classified as open-weight, meaning their parameters are available for public access and can be operated independently or through third-party platforms like OpenRouter. This accessibility provides developers with flexibility, eliminating the need for deep negotiations or proprietary contracts. It is as simple as updating a base URL and an API key to switch to a different model.
Implications for the AI Market When Chinese alternatives suddenly dominate over half of the top-model token volume on a significant platform in just a week, it sends a clear message. The performance of these models is now comparable enough for most operational workloads, making cost the primary factor for developers and businesses.