As the focus of AI competition shifts from model performance to token pricing, a survey has found China is starting to pull ahead in the fight.

According to a Financial Times report on March 26, OpenRouter data tracking AI model usage showed that from February, token consumption by Chinese AI models such as DeepSeek and MiniMax overtook that of U.S. competitors. Token use for MiniMax's M2.5 model surged 476 percent as of March 20 from a month earlier.

Tokens, the units of text, code and data processed by AI models, have emerged as an important metric in AI, to the extent that Nvidia CEO Jensen Huang has called them "the key that will drive the AI economy". Because developers pay in proportion to token usage, tokens are seen as a measure that can determine model adoption rates.

At present, token prices vary widely across providers.

Chinese companies such as MiniMax and Moonshot charge $2 to $3 per 1 million output tokens. Anthropic's Claude Sonnet 4.5 costs about $15, a gap of nearly sixfold.

The FT cited Hong Kong developer Terry Zhang as saying, "I used to use only Claude, but now I use Moonshot's Kimi model for 80 percent of my work and cut daily spending to $50. If I use only Claude, it comes to $900 a day."

As AI spreads, the price gap could widen further.

If a simple chatbot uses 30,000 tokens to summarise Shakespeare's Hamlet, an AI agent consumes up to 20,000,000 tokens for a basic coding task. Will Liang, CEO of Sydney-based technology consultancy Amplify AI Group, said, "If an agent consumes millions of tokens a day, even a small difference in price per token takes up a big share of costs. The more agent adoption grows, the more the structure favours Chinese companies." China may not be able to maintain its edge in the competition. Zhipu AI's GLM-5 model rose into OpenRouter's top ranks in February, but experienced service disruptions after failing to handle surging demand, and after an apology and a price hike its shares fell 22 percent in a day, wiping out more than $10 billion in market value.