AI agents now consuming more tokens than humans on OpenRouter, up 14x since February

The shift toward AI-to-AI consumption reflects a fundamental change in how AI systems are deployed. Rather than directly serving human queries, agents now autonomously spawn additional AI processes to accomplish tasks, creating a multiplication effect in token demand. This pattern has emerged as models become more capable of independent reasoning and planning.

The disparity between token growth and cost growth reveals an important nuance: while raw token consumption by agents accelerated 14x versus 2.8x for humans, the actual economic impact is dampened by cached prompts. Since February 2026, cached prompts have accounted for nearly 70 percent of agent token usage, and these incur substantially lower billing rates. This suggests that while AI systems are consuming vastly more computational resources in absolute terms, the marginal cost to AI providers may be lower than headline token figures imply.

OpenRouter's data may actually understate the scale of this shift among major AI labs. OpenRouter predominantly hosts open-weight models, which are inherently less token-efficient than proprietary models from OpenAI or Anthropic. If the trend mirrors the broader industry, large-scale AI systems at major providers could be experiencing similar or even more pronounced agent-driven token inflation, partly driven by reasoning models that deliberate extensively before responding.