Chinese AI model weekly token usage leads the world for the 20th consecutive week, Hunyuan Hy4 preview ranks second globally

On September 14, 2026, OpenRouter data published for the week of September 7-13 showed global AI model API traffic at 127 trillion tokens, up 10.43 percent week-over-week. Chinese AI models logged 61.17 trillion tokens of weekly traffic, up 7.85 percent week-over-week, versus 21.76 trillion tokens for US AI models (up 31.56 percent), marking the twentieth consecutive week that Chinese-vendor traffic exceeded US-vendor traffic. Four of the top five models globally by weekly token usage were from Chinese vendors: Tencent Hunyuan Hy4 preview ranked second at 16.8 trillion tokens (up 15 percent week-over-week), Zhipu GLM 5.3 Flash third at 11.9 trillion tokens, and DeepSeek V4 Flash 0731 fourth at 11.6 trillion tokens. Xiaomi MiMo-V2.5 rose to fifth at 7.77 trillion tokens (up 230 percent week-over-week), while DeepSeek V4.1 Flash, released September 10, climbed to sixth within three days of launch at 4.94 trillion tokens. Chinese vendors are no longer fighting with single models but with portfolios: Tencent runs Hy4 preview and Hy3, DeepSeek runs V4 Flash 0731, V4 Flash 0423, and V4 Pro 0423, Zhipu runs GLM 5.3 Flash alongside Ox Alpha and other lines; Chinese models now occupy more than seven of the top ten spots on the leaderboard, accounting for about 71.6 percent of aggregate traffic. The numbers land the same day as Beijing's official response to Amodei's essay, and have been read as substantive evidence that China's AI stack continues to lead at the application and developer-ecosystem layer despite ongoing chip and equipment export controls.