OpenRouter data shows that AI agents have consumed more tokens than humans since February 6, 2025, with agentic usage growing 14x while human usage grew only 2.8x. However, nearly 70% of agent token consumption comes from cheap cached prompts, so actual costs are rising far more slowly than the raw numbers suggest.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
OpenRouter data shows that since February 6, 2026, token consumption by AI agents has grown 14-fold, while human usage has only grown 2.8-fold. This marks a point where AI-to-AI interactions may have surpassed human usage, but nearly 70% of agent tokens come from cached prompts, so actual cost growth may be lower than the raw numbers suggest.
According to OpenRouter analyst Peter Walker, February 6, 2026, may have been the last day humans consumed more tokens than AI agents. Since then, agent token usage has grown 14-fold, while human usage has grown only 2.8-fold. This data from the OpenRouter platform reflects the trend of AI agents consuming large amounts of tokens during autonomous work.
Specifically, since February 2026, AI agent token consumption on OpenRouter has grown from 0.51 trillion to 7.3 trillion, an increase of about 14-fold. Human usage grew only 2.8-fold over the same period. These figures come from OpenRouter's public data, provided by analyst Peter Walker.
Despite the surge in agent token usage, nearly 70% of agent token usage comes from cached prompts, which are billed at lower rates. This means actual cost increases may be slower than the raw numbers suggest. Cached prompts allow reuse of previously computed results, reducing the cost per call.
This finding is crucial for understanding the economic impact of AI agents. While token volume growth is striking, cost growth may be more moderate. However, the raw numbers still reflect a substantial increase in AI agent activity.
The OpenRouter platform skews toward open-weight models, which are typically less token-efficient than models from OpenAI or Anthropic. Therefore, the growth on OpenRouter may partly reflect differences in model selection. However, analysts believe the trend may be similar across major labs.
Reasoning models have already caused token inflation, as they think longer before responding, even when not necessary. This suggests that token consumption growth may partly stem from changes in model design, not just an increase in the number of agents.
This report is based on public data from OpenRouter analyst Peter Walker and is attributed to the source. Specific figures (such as 14-fold, 2.8-fold, 70%) come from the analyst and have not been independently verified. OpenRouter's data may skew toward open-weight models, so the trend may not fully represent the entire industry.
AI agents are becoming the primary consumers of AI services, but cost growth may be tempered by caching discounts. This trend could reshape the business models of the AI industry, but more data is needed to verify its universality.
Primary report
Primary source