AI Agents Overtake Humans as Top Token Users on OpenRouter

What happened

AI agents have overtaken humans as the largest consumers of tokens on OpenRouter, the AI model routing platform, according to data reported by The Decoder. Agentic token usage has grown 14-fold since 6 February 2025, the point at which agent consumption first exceeded human consumption on the platform, while human-driven usage rose by a comparatively modest 2.8 times over the same period.

The headline growth figure masks a more measured cost story: nearly 70 percent of the tokens consumed by agents come from cached prompts, which are significantly cheaper to serve than freshly processed requests. As a result, the actual computing and billing burden associated with agentic traffic is rising far more slowly than the raw 14x usage figure implies.

Why it matters

This is a clear, quantified signal that autonomous AI agents — software that plans, calls tools and executes multi-step tasks with minimal human prompting — are becoming a distinct and dominant class of AI consumer in their own right, separate from human chat-based usage. For organisations building or deploying AI infrastructure, it points to a shift in what "demand" for AI capacity actually looks like: increasingly repetitive, machine-to-machine and cache-friendly, rather than the more varied, one-off queries typical of human users.

For platform operators and enterprises budgeting for AI adoption, the caching detail is the more strategically important finding. It suggests that agentic workloads, if architected well, do not need to scale costs linearly with token volume — creating room to expand agent deployment without a proportional rise in spend, provided caching and prompt-reuse strategies are built in from the start.

By the numbers

• 14x growth in agentic token usage on OpenRouter since 6 February 2025.
• 2.8x growth in human token usage over the same period.
• 70 percent of agent token consumption comes from cached prompts.

The Renascence take

Most coverage of this data will fixate on the 14x growth number as evidence of an agent boom. The more useful signal for operators is the caching statistic, because it reveals how agentic systems actually behave in production — not as novel, exploratory users but as repetitive, procedural ones.

Agents behave nothing like humans as "customers" of AI infrastructure, and treating them the same way in cost models or capacity planning is a mistake. The 70 percent cache figure tells you agentic workloads are structurally repetitive — which is good news for cost control, but also a warning sign if those repeated prompts are masking inefficient or poorly designed agent logic rather than genuine reuse. Before scaling agent deployments, experience and platform teams should audit what's actually being cached: is it smart, deliberate prompt engineering, or is it a symptom of agents looping through the same steps because the underlying workflow hasn't been designed with efficiency in mind? The organisations that win here will be the ones that treat agent-token economics as a service-design problem, not just an infrastructure bill.