China's AI Models Top OpenRouter Rankings for 15 Straight Weeks

OpenRouter's latest weekly token rankings show global AI model traffic reached 69 trillion tokens between August 3 and August 9, up 21.48% from the previous week. Chinese models accounted for 34.25 trillion tokens, up 21.76% week over week, while US models logged 9.17 trillion, up 109.36%. That means Chinese model traffic has now topped the US on OpenRouter for 15 consecutive weeks.

All four of the week's most-called models came from China: DeepSeek-V4-Flash-0731, the just-released official version, led with 8.83 trillion tokens, up 570% week over week; Tencent's Hy3 followed with 8.05 trillion (up 67%); the DeepSeek-V4-Flash-0423 preview ranked third with 5.88 trillion (down 19%); and Xiaomi's MiMo-V2.5 took fourth with 5.39 trillion (down 14%).

The data quickly drew attention, with some reading it as evidence that "China's AI has overtaken the US." Citing an FT Chinese analysis, the report notes that AI compute spending is shifting from model training to large-scale inference, and that the rise of agents is moving applications from Q&A to complex task execution, driving token consumption sharply higher. Cost is where Chinese models excel: prices run as low as 18 US cents per million tokens versus about $4 for US models — a more than 20-fold gap.

The report cautions, however, that topping the call rankings is not the same as "China's AI surpassing the US." OpenRouter is an aggregation platform, and its data covers only traffic routed through its gateway by third-party developers, overseas users and downstream applications — excluding vendors' own apps, official sites, consumer products, internal systems and private deployments.

Even with that caveat, the numbers carry weight. More US companies are adopting Chinese models: AI startup Lindy switched entirely from Anthropic's Claude to DeepSeek-V4 in June after its API bill exceeded its total payroll, cutting inference costs by roughly 95%; and Zhipu's GLM-5.2 saw daily token usage grow about 27-fold and customer counts about 80-fold within a week of release on Vercel's gateway.

Goldman Sachs chief China equity strategist Liu Jinjin said AI tokens could become a major new export growth engine for China, while Brookings researcher Kyle Chan estimates Chinese models are "close to US frontier models," with a performance gap of roughly six to nine months.

The next weekly ranking will show whether the US' 109% growth rate can sustain its catch-up and which Chinese releases pick up where DeepSeek-V4-Flash's surge leaves off. Either way, the competitiveness of Chinese models on openness and cost is now visible in real usage data.