OpenRouter token traffic jumps 25,000% since January as reasoning models and agents inflate usage

Weekly token consumption on OpenRouter has climbed from 0.5 trillion in January 2025 to 126.2 trillion, a rise of more than 25,000 percent in roughly a year. The chart tracking that curve has become a shorthand for the argument over whether AI demand is real or inflated. The number is accurate, but it measures throughput on one routing platform, not the number of people or businesses actually getting value from AI.

OpenRouter is a gateway that lets developers send requests to many different models through a single API, so its traffic aggregates usage across providers rather than representing one company's customer base. That makes it a useful proxy for developer activity and a poor proxy for end-user demand, because the same workload can generate wildly different token counts depending on which model is called and how the request is structured.

The biggest single driver of the increase is the shift toward reasoning models, which generate long internal chains of thought before producing a final answer. A question that once cost a few hundred tokens can now consume thousands, and those hidden tokens are billed and counted even though the user never sees them. Reasoning has become a default setting across major model families, so the same tasks that ran cheaply a year ago now run expensively by design.

A second driver is the spread of AI agents: systems that loop through tool calls, retries, and self-checks with little human oversight. Many of these agents are built without token budgeting, meaning they keep re-reading context, re-planning, and re-querying until a task completes or a limit is hit. A single agent session can burn through millions of tokens, and developers running fleets of them can move aggregate numbers on their own.

That distinction matters for the bubble debate. Token volume growth is often cited as proof that AI adoption is compounding, but volume can rise while the number of paying users stays flat, if each user's workloads simply get heavier. Falling per-token prices also mean revenue does not scale with the token curve, so a 25,000 percent jump in consumption can coexist with much more modest growth in spend.

What to watch next is whether token growth starts tracking user and revenue growth, or keeps diverging from it. OpenRouter's own data cannot settle that question, and neither can a single chart. The clearer signal will come from providers that disclose paying customers, retention, and gross margins alongside raw usage, and from whether the next generation of models ships with cheaper reasoning rather than longer reasoning.