Peter Walker noted on 6 February 2026 that AI agents may have surpassed humans as the largest consumer of tokens on OpenRouter. Agentic usage has grown fourteen times since that date, while human consumption has only increased by two point eight times. These systems now operate independently for extended periods, spawning additional AI processes to complete tasks. However, nearly seventy percent of this volume consists of cached prompts, which incur significantly lower billing rates than fresh queries. Consequently, the financial impact on providers is less severe than the raw token figures initially suggest.
The shift indicates a move toward automated workflows where machines coordinate complex chains of reasoning without direct human intervention. OpenRouter hosts open-weight models that typically require more tokens than those from OpenAI or Anthropic, yet the trend likely mirrors activity at major labs. Token inflation began earlier with reasoning models, which engage in prolonged internal deliberation before generating responses. This behaviour inflates usage metrics even when the model does not need to think that extensively.
* Cached prompts account for nearly seventy percent of agent volume
* OpenRouter skews toward open-weight models
* Reasoning models drive inflation through extended deliberation
