China's daily AI token usage hits 500 trillion as model speeds double

China's top telecom research body CAICT has released its 2026 report on large model token services. As AI models evolve fast, China's daily token use hit about 500 trillion in June, nearly 4 times the level of December 2025.

In August, Chinese models produced an average of 78.23 tokens per second, double the 2025 figure, but still behind top foreign models at 112.82. The wait time for the first token rose to 1.26 seconds, as models got bigger and usage grew.

On price, Chinese model services averaged 13.46 yuan in 2026, nearly double from the second half of 2025, while foreign models cost about 117.8 yuan. Models with a 1 million-token context window now make up 47.8%, and those with 200K or more reach 95.7%.

The open-source models DeepSeek, Qwen, GLM and Kimi have improved a lot. The same model can run quite differently across platforms, with speed gaps of up to nearly double, mainly due to different optimization strategies.