AI agents now burn more tokens than humans on OpenRouter
OpenRouter's own usage data marks a quiet milestone in the agentic shift: since February, AI agents have consumed more tokens on the platform than the humans who set them loose — and agentic demand has grown about 14x while human usage is up just 2.8x.
AI agents now consume more tokens than humans on OpenRouter — agentic usage is up roughly 14x since February. OpenRouter analyst Peter Walker walked the numbers on LinkedIn, and the platform's own data backs it: agent token demand has climbed from about 0.51 trillion tokens a month to 7.3 trillion, while human-driven usage grew only 2.8x over the same stretch. February 6 may have been the last day individual users out-burned agents. The trend is far from aesthetic: agents aren't just chatting — they spin up sub-agents, run long tool loops, and chew through context windows for hours at a time, so each active agent multiplies its own consumption.
The headline number overstates the real cost spike, and that's the detail worth holding onto. Nearly 70 percent of agent token consumption comes from cached prompts, which are billed at much lower rates than fresh reasoning tokens. So AI is paying AI's bills less expensively than the raw volume would suggest — actual dollars are climbing far more slowly than the raw 14x figure implies. There's a compounding effect at work here, too. OpenRouter skews toward open-weight models, which are generally less token-efficient than their frontier peers, and reasoning models already inflated token spend by "thinking" before they answer.
Why it matters: when agents start consuming more tokens than the people driving them, the economic center of gravity moves. Token inflation already turned tokens into the field's core business metric — this data shows the largest buyers of inference are now the agents themselves, not end users. OpenRouter's mix isn't identical to the closed labs, but the direction is the story: if agents are the biggest demand source, pricing, supply planning, and even which models get optimized all start to follow the machine's appetite instead of the human's. That's a structural shift in who the AI business is actually selling to.
What to watch: where margin compression shows up first. If 70 percent of agent tokens are cheap cached ones, providers have room to absorb demand — but the moment agents start needing fresh, expensive reasoning tokens in volume, that cushion thins fast and latency-sensitive workloads get the squeeze.
Are we approaching the point where AI's biggest end-customer is other AI? Tell us in the comments.
