China: AI Token Prices Double and Usage Triples

A report issued by the China Academy of Information and Communications Technology stated that prices for AI model token services in China have doubled since the second half of 2025, while usage volume has tripled. Average daily token usage reached about 500 trillion in June 2026, nearly four times its level in December 2025.

The report explained that the average domestic price rose from 6.8 yuan (about $1.02) in the second half of 2025 to 13.46 yuan this year, after adjusting billing methods for comparison. The domestic price remains significantly lower than overseas prices, as foreign models launched since 2026 average about 117.8 yuan on the same basis, nearly 8.7 times the domestic level.

Wang Peng, a researcher at the Beijing Academy of Social Sciences, said that the combination of rising prices and increased usage indicates that demand for models has become strong enough to absorb a price recovery, and that the market is shifting toward competition based on service quality. He added that the competitive logic has shifted from burning cash to attract customers to earning through service.

The market witnessed the launch of new models, including 'Kimi K3.1 Pro' from Moonshot AI and models from ByteDance's Volcano Engine unit. Volcano Engine has also begun enterprise trials for a 4K video generation model with 20 studios and directors. Additionally, a national AI application center in the automotive sector announced a core platform managing 112.7 exaflops of computing power.

What do these terms mean?

Token: A small unit of text processed and billed by the model, which can be a word, part of a word, or punctuation mark.

Model-as-a-Service: Selling access to an AI model via the cloud with pay-as-you-go pricing, instead of purchasing and running the model internally.

API: A technical channel that allows a software program to send requests to the model and automatically receive its responses.

Exaflop: A unit for measuring supercomputer speed, equal to about one quintillion floating-point operations per second.