Chinese models draw 4.5x US token usage on OpenRouter as all top five slots go to China

A report published on October 18 by Chinese outlet Guoshi Zhitongche (国是直通车), citing data from model marketplace OpenRouter, puts the week of October 5-11 at 172 trillion tokens processed globally. Chinese models accounted for 86.74 trillion of that total, against 18.97 trillion for US models — roughly 4.5 times as much — and the five most-used models in the world that week were all Chinese.

Token volume is a rough but useful proxy for real usage. Every API call represents a concrete interaction, whether a customer-service bot, a code completion, a marketing draft or a data analysis job, so the metric measures how much models are actually used rather than how much they are discussed. According to OpenRouter, China's weekly token volume had stayed ahead of the US for 24 consecutive weeks as of October 11.

Price is the most direct driver. The report cites Justin Summerville, who works on data and analytics at OpenRouter, telling CNBC that Chinese open models are typically 60% to 90% cheaper than the leading Anthropic and OpenAI models. For teams running high-volume inference, that is not a marginal discount; it is the difference between a feature that stays in pilot and one that ships.

Open weights and local-language tuning turn that price advantage into deployments. In Southeast Asia, where more than 1,200 languages are spoken, the report says mainstream Western models were trained on data that was over 80% standard English, with Southeast Asian languages at one point accounting for only 0.5%. Qwen and other Chinese models covered hundreds of local languages during pretraining, according to the report, and Singapore's national 'Sea Lion' model has moved onto an Alibaba Qwen base.

In the US and Europe, procurement is increasingly a cost exercise. The report cites the Financial Times saying DoorDash, Airbnb and Siemens are among large companies turning to Chinese AI models, and cites US investment-firm data suggesting as many as 80% of American AI startups use Chinese open models in fundraising pitches. In Kenya and Nigeria, developers use downloadable weights as a local toolkit for education tools, chatbots and business-process automation.

Performance is catching up, which makes the choice less purely about price. The report quotes Stanford's 2026 AI Index as saying US and Chinese models have traded the top spot repeatedly since early 2025, with the gap between the best models at just 2.7% as of March 2026. Intel CEO Lip-Bu Tan is quoted saying Chinese frontier models are not only cheaper but 'good enough' for real applications, while Nvidia CEO Jensen Huang has repeatedly said open Chinese models lower the barrier to AI use.

The advantage is still concentrated in cost-sensitive, high-volume workloads. CNBC's piece, as summarized by the report, argues the US retains a lead in coding and agentic capabilities, but that Chinese models — downloadable, modifiable and deployable on-premises — are increasingly likely to become the default choice globally, especially in developing economies.

The question now is durability. Weekly token counts move with promotional pricing, free tiers and a single viral app, and 24 weeks of leadership describes a trend rather than a proven capability lead. The harder test for Chinese vendors is moving from being called to being depended on: whether enterprises entrust core workflows, data pipelines and compliance obligations to these models, and whether the open ecosystem keeps iterating once performance parity is reached.