Data released on October 8th by OpenRouter, an AI model aggregation platform, shows that the token consumption of AI agents on the platform has surpassed that of human users across the board. On February 6th, agents first overtook humans in token consumption; by the last week of September, agent-type API keys consumed approximately seven times more tokens than human-type keys, and OpenRouter expects this gap to widen further in the coming months. The platform categorizes API keys monthly into "Agent-type," "Human-type," or "Hybrid-type" based on characteristics such as tool call frequency, number of rounds per dialogue, and request sending intervals.
The usage structure of agents differs significantly from that of humans. Because the complete context must be resent after each step, nearly 90% of agent usage in September was cache reads, compared to approximately 60% in January, with output accounting for only about 1%. The cost of cache reads is about one-tenth of new input tokens; using Claude Sonnet 5 as an example, the cost per million agent tokens is approximately $0.51, compared to approximately $2.08 without caching. Analysts believe that the computing power requirements of large models are rapidly shifting from human-computer dialogue to automated intelligent agent workflows, which will reshape the demand structure of model manufacturers and the computing power supply chain.
