Chinese AI Models Overtake US Rivals in Global Token Consumption
They are no longer called GPT or Claude, but GLM, MiniMax, or Kimi: Since February 2026, Chinese AI models have recorded more token consumption on the OpenRouter platform than their American competitors. This shift is no coincidence, but the result of deliberate strategies around pricing, efficiency, and the growing demand for autonomous AI agents.
What is OpenRouter?
OpenRouter is the world’s largest aggregation platform for large language models (LLMs). The company was founded by Alex Atallah, the former CTO of the NFT platform OpenSea. Through a unified programming interface (API), OpenRouter offers developers and companies access to more than 400 AI models from over 60 providers worldwide.
The platform functions as a kind of intermediary: instead of signing up directly with OpenAI, Anthropic, or a Chinese provider, users can access and compare all models via OpenRouter. Because OpenRouter publishes detailed usage data, the platform is regarded in the industry as an important indicator of global AI trends, even though it only reflects a portion of worldwide model consumption.
What is a token and why does it matter?
A token is the smallest unit processed by an AI language model. It can be an entire word, part of a word, a punctuation mark, or a single character. AI providers typically charge for their services based on the number of tokens processed, both for incoming requests (input) and generated responses (output). Token consumption is therefore simultaneously a measure of a model’s usage intensity and a central pricing battleground.
Nvidia CEO Jensen Huang recently emphasized that the production and use of tokens will drive the AI economy of the future. This becomes particularly relevant with the rise of so-called AI agents: while a simple chatbot requires around 30,000 tokens for a summary of Shakespeare’s “Hamlet,” an AI agent can consume up to 20 million tokens for a smaller programming task.
The numbers: Chinese models dominate the rankings
According to OpenRouter data from February and March 2026, Chinese models account for around 61 percent of total token consumption among the ten most-used models on the platform. In the week of March 16–22, Chinese models recorded a total of 7.36 trillion tokens, an increase of 56.9 percent compared to the previous week.
| Model | Provider | Weekly token usage | Change |
|---|---|---|---|
| MiniMax M2.5 | MiniMax (China) | 2.45 trillion | +197% |
| Kimi K2.5 | Moonshot AI (China) | 1.21 trillion | -20% |
| GLM-5 | Zhipu AI (China) | 780 billion | +158% |
| DeepSeek V3.2 | DeepSeek (China) | n/a | Rank 5 |
Total weekly token consumption on OpenRouter rose to 12.1 trillion, a 12.7-fold increase compared to the previous year. Chinese models accounted for 5.3 of the 8.7 trillion tokens consumed by the top 10 models.
The decisive factor: Price
A central reason for the rise of Chinese models is their considerable cost advantage. The price differences between Chinese and American providers are substantial:
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| MiniMax M2.5 | $0.30 | $1.10 |
| Moonshot Kimi | approx. $2 to $3 | approx. $2 to $3 |
| Anthropic Claude Opus 4.6 | $5.00 | $25.00 |
| Anthropic Claude Sonnet 4.5 | n/a | approx. $15.00 |
Chinese models are therefore 10 to 20 times cheaper than leading American alternatives, depending on the comparison. This difference is particularly noticeable with AI agents, which continuously process large volumes of tokens.
Why are Chinese models so affordable?
The cost advantage of Chinese AI companies has several causes. First, they benefit from cheaper energy, partly through massive state investment in renewable energies. The Chinese government has explicitly enshrined the linking of energy policy and AI competitiveness as a national priority in its 2026 work plan.
Second, Chinese companies have developed more efficient AI architectures, such as so-called “Mixture-of-Experts” designs, which reduce computational overhead. This focus on efficiency was also driven by US export restrictions on advanced chips, which forced Chinese developers to achieve more with less computing power.
Third, China’s LLMs are simply available via APIs at very low prices.
AI agents as growth drivers
The boom in Chinese models coincides with a fundamental shift in AI usage. Programming activities now account for more than half of total token consumption on OpenRouter, compared to just 11 percent at the start of 2025. At the same time, AI agents — which autonomously execute multi-step tasks — now generate the majority of output tokens.
Even among US start-ups, the adoption of Chinese models is considerable: venture capitalist Martin Casado of Andreessen Horowitz estimates that around 80 percent of young AI companies using open-source stacks now use Chinese models.
Performance on par with the competition
The price difference alone does not fully explain the success. Chinese models have made significant technical progress. MiniMax M2.5 achieved a score of 80.2 percent on the SWE-Bench Verified benchmark, a standard test for programming capabilities. Anthropic’s Claude Opus 4.6 is only slightly ahead at 80.8 percent. According to the analytics platform Artificial Analysis, several Chinese models have reached the global top tier, and the gap to leading international models in areas such as programming, multimodality, and long-text processing continues to shrink.
Limitations and risks
The rise of Chinese models has not been without setbacks. Zhipu AI’s GLM-5 briefly topped the OpenRouter charts in February before demand exceeded available computing capacity. Delays and service outages occurred, the company had to issue a public apology and raise prices. The stock subsequently lost 22 percent in a single day, equivalent to a market value loss of more than 10 billion US dollars.
