The customer base for AI has officially flipped. Peter Walker, an analyst at OpenRouter, pegs February 6, 2026 as the last day human users consumed more tokens than AI agents did. Since then, agentic token usage has exploded 14x, while human usage has grown a comparatively modest 2.8x. That’s not a gradual shift — it’s a crossing of the Rubicon, and the agents aren’t looking back.
What’s driving the surge? Agents increasingly operate autonomously over extended periods, spinning up additional AI processes as they work. One agent calls another, which calls another, and the token meter keeps spinning. OpenRouter’s numbers skew toward open-weight models, which tend to be less token-efficient than the proprietary systems from OpenAI or Anthropic, so the raw figures may overstate the trend somewhat. But the direction is unmistakable, and the major labs are likely seeing a similar pattern. Token inflation was already underway with reasoning models, which burn through tokens thinking longer before responding — sometimes whether they need to or not.
Here’s the wrinkle that should temper any panic about runaway infrastructure costs: nearly 70 percent of agent token usage comes from cached prompts. Cached tokens bill at substantially lower rates than fresh ones, so the actual dollar cost of all this agent-to-agent chatter isn’t climbing nearly as fast as the raw token counts suggest. The economy of AI consumption is decoupling from the volume of AI consumption.
Still, there’s a question worth sitting with. If agents are the primary consumers of AI output, what does that mean for the models themselves? Optimizing for machine-readable responses is different from optimizing for human-readable ones. The feedback loop is changing — and the products built on top of that loop will have to change with it.
