The Era When AI Surpasses Humans in AI Usage
Since February 6, 2025, OpenRouter, an API gateway that consolidates multiple AI models, has witnessed AI agent token consumption exceeding human consumption, subsequently expanding 14-fold. Meanwhile, human usage growth remained at 2.8x, with agent-type AI usage where AI invokes other AI rapidly becoming mainstream. Approximately 70% of token consumption is handled through cost-effective caching, meaning actual cost increases are less dramatic than the raw numbers suggest.
An era in which AI uses more AI than humans is quietly beginning. OpenRouter, an API gateway that consolidates multiple AI models, has witnessed a turning point since February 6, 2025, where token consumption by AI agents exceeds that of humans. A token is the smallest unit by which AI processes text, and the volume of consumption directly indicates the scale of AI usage.
The subsequent growth trajectory is starkly contrasting. Token consumption by AI agents expanded approximately 14-fold after February 6. In contrast, the increase in token usage by humans remained at 2.8x, with the gap between the two continuing to widen. This figure demonstrates that not only are individuals and organizations using AI as a tool, but 'agent-type' usage—where AI invokes other AI to advance work—is rapidly becoming mainstream.
However, the pace of cost increase is not as dramatic as consumption growth. Approximately 70% of the tokens consumed by agents are processed through a mechanism called 'cached prompts.' This is a feature that allows tokens to be processed more economically than usual by reusing prompts (instructions) that have been used before. Consequently, even as usage volume increases 14-fold, actual cost increases occur at a much more gradual pace.
The backdrop involves the broader proliferation of AI agents. In recent years, the development and implementation of 'agent-type AI'—where multiple AIs collaborate to autonomously perform complex tasks over extended periods—has been accelerating. Traditional AI usage centered on humans providing instructions each time, but in agent-type usage, once a goal is provided, AI independently judges and executes tasks through multiple steps. In such mechanisms, a single operation consumes vast quantities of tokens, creating a structure where consumption readily expands dramatically.
This shift is significant for the industry because it is fundamentally rewriting the demand structure of AI infrastructure itself. Until now, demand for AI services has been primarily defined by 'how many times humans use it,' but with the rise of agent-type usage, a new dimension of 'how many times AI calls other AI' has been added. Since this dimension operates independently of human behavioral pace, both demand forecasting and infrastructure design are understood to diverge substantially from conventional approaches.
Although caching functions serve as a buffer on the cost front, should agent-type usage expand further, the impact on infrastructure costs will become unavoidable. Going forward, how efficiently tokens can be consumed while executing sophisticated tasks will emerge as a critical factor determining the competitiveness of AI agents. We are entering a phase where not only the 'quantity' of usage, but also the 'quality and cost efficiency' are called into question.
This article is an original work independently written and edited by the AI issue editorial team based on factual reporting. © AI issue. Unauthorized reproduction, redistribution, or use for AI training is prohibited.