AI agents are operating independently, performing complex sequences of operations in autonomous mode. They access various services and databases and verify their own results, which requires the use of tokens. According to OpenRouter, agentic requests generate up to 15 times more tokens than standard human-created requests.
By August, agent activity on the OpenRouter platform had exceeded human activity by more than five times. Token consumption on the platform grew 14-fold by the beginning of August compared to six months prior. TH Capital experts estimate that the use of tokens by AI models will continue to grow exponentially.
In China, daily token usage grew from 100 billion in early 2024 to 500 trillion by June of the current year, a 5,000-fold increase. This surge in demand leads users to distribute less critical workloads to lower-cost models to avoid the high costs of advanced systems.
DeepSeek V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. In contrast, OpenAI's GPT-5.6 Sol costs $4 per million input tokens and $20 per million output tokens, while Anthropic's Fable 5.1 costs $10 and $50 respectively.
OpenRouter statistics show that DeepSeek solutions are the most active with a 22.7% share of requests, compared to 20.4% for Google and 19.6% for OpenAI. This trend suggests a shift in favor of cheaper Chinese AI models as token consumption increases.
