AI Is Now AI's Biggest Customer: Agent Token Usage on OpenRouter Grew 14x

AI agents on OpenRouter now consume more tokens than human users, and the gap is widening fast: agent consumption grew 14x, while human consumption grew just 2.8x. The Decoder breaks down the statistics from the largest language model aggregator.

Machines Have Overtaken Humans

OpenRouter is a handy thermometer for the entire AI industry: through a single API, the platform routes requests to hundreds of models from dozens of providers, and its public statistics reflect real usage patterns rather than marketing claims. The thermometer shows a structural shift: the point where agent traffic exceeded human traffic has already been passed, and the curves have only been diverging since.

Behind the dry numbers is a change in the very nature of AI consumption. A human asks a question and reads the answer: a handful of requests per minute. An agent working on a task runs in loops: it plans, calls tools, reads the results, and checks its own work, and every turn of the loop burns tokens. A single working agent generates the traffic of dozens of people.

• Agent token consumption grew 14x, human consumption 2.8x
• Agents consistently consume more than humans, and the gap is growing
• Nearly 70% of agent traffic is cached prompts at a reduced price

Why Costs Are Growing Slower Than Traffic

The Decoder points to an important caveat: about 70% of agent consumption is cached prompts. At every step, agents send the model the same context: system instructions, task history, documentation. Providers charge several times less for reprocessing such fragments. As a result, actual costs are growing much more slowly than the raw token counter.

For model providers, this is a double-edged piece of news. Explosive agent demand confirms the market. But the structure of that demand, dominated by cheap cached tokens, means revenue scales worse than traffic, while the infrastructure load remains real.

What This Means

The AI economy is shifting from chats to pipelines: the main consumer of models is no longer a human with a question but a program with a task. That changes product requirements. Latency, caching, and the cost of a loop matter more than a polished interface, which explains why providers are racing to cut prices specifically on cached and batch requests. The next phase of model competition will be won on the economics of the agent loop, not on benchmark scores.

Frequently Asked Questions

What is OpenRouter?

An aggregator for access to language models: a single API through which developers reach hundreds of models from different providers. The platform's public statistics have become a popular indicator of real demand for models.

Why are cached prompts cheaper?

Providers store context fragments they have already processed and do not recompute them when they are sent again. That means less computation, and the rate for such tokens is several times lower than the standard rate.