The Maze: Agents have become the dominant source of token traffic in an OpenRouter exhibit published by Andreessen Horowitz. OpenRouter connects developers to AI models through one interface. Agentic usage reaches a seven-day average of 7.3 trillion tokens, with a labeled fourteenfold increase since February. That is a sharp change in what consumes computing capacity. It says much less about how many people use AI, how much they pay, or how much useful work gets finished.
The crossover came before the explosion. Agentic token usage overtook the Human category on February 6, 2026. Before that, both remained small beside the later endpoint. Through spring and summer, agentic traffic pulled away while Human and Mixed usage stayed far below it. The important shift is the composition of demand on this platform. A person can start an automated workflow that makes many model calls, so a smaller audience can generate a much larger share of activity. More tokens therefore need not mean more customers.
The rise was steep and uneven. Agentic usage shows an April spike and reversal, a strong June acceleration, and another late-summer peak and dip before reaching 7.3 trillion tokens. Human and Mixed also increase, but do not follow the same scale of expansion. The source labels the agentic increase as 14x since February. Intermediate positions are approximate visual readings, not recovered daily observations. The swings matter for capacity planning, but the evidence does not identify which products, releases or customer groups caused them.
The headline has a denominator problem. The post describes agents consuming nearly five times the tokens people do. That compares aggregate usage categories, not an agent and a person completing the same task. OpenRouter's methodology counts prompt and completion tokens in its model rankings and excludes private requests. Those rankings measure routed traffic, not users or spend. The historical Agentic, Human and Mixed classification rules were not recovered here, so treat the categories as publisher labels rather than a universal census of AI demand.
Volume is a billable activity, not a business outcome. Tokens are pieces of text processed by a model, and their volume can rise as automated workflows make repeated requests. A retailer using agents to inspect product listings or handle service cases could see more model traffic even with stable customer demand. That is a possible operating mechanism, not an explanation proven by this dataset. Higher usage can reflect more completed work, longer instructions, extra checking or unsuccessful loops. The exhibit does not separate those causes or measure their financial returns.
The operator's next metric belongs outside the token counter. Keep traffic visible, but pair it with accepted outputs and the full cost of producing them. For a catalog workflow, that could mean listings corrected to the required standard; for service, cases resolved without human rework. Count failed attempts and review effort alongside successful runs. These are practical measurement choices, not performance results from the source. A growing usage total is useful infrastructure evidence. It becomes commercial evidence only when someone connects that activity to a finished task and a defensible cost.
Why it matters: An agent can turn one human instruction into a stream of AI requests. This OpenRouter snapshot makes the scale of that shift hard to ignore, while leaving its economics unresolved. Ecommerce operators should watch workload composition and budget limits as they automate repeatable work. The useful question is how much accepted output the spend buys. Token growth can justify more capacity; it cannot, on its own, justify a productivity claim or prove that the automation has earned its place.
Images: Cover AI-generated
