Xianzi Insights

OpenRouter's Numbers: AI Agents Now Use 5× More Tokens Than Humans — What 7.3 Trillion Tells Us

林政賢 · 2026-08-27

You open ChatGPT, type an instruction, wait for the answer, and think that's the whole of "using AI". Input, wait, output — you're the protagonist, AI is the tool. That picture isn't wrong. It just isn't the whole picture any more.

You think you're using AI. Actually, AI is using AI

OpenRouter is a platform that puts hundreds of AI models behind one entrance, so it sees the traffic picture nobody else can. The figures it published in August 2026 look like this:

A token is the billing unit for text an AI processes — think of it as "fuel burned while the AI thinks". So while you send one prompt and get one reply back, five times that volume is AI issuing instructions to other AIs, waiting for their answers, and moving to the next step.

The scene you imagine is "human asks, AI answers". Most of it has already become "AI asks, AI answers", and you're just the one who pressed the first button.

From "tool" to "user": the role of AI is changing in kind

AI usage used to mean "how many people are asking questions". You type a prompt, the model returns an answer, back and forth, with a human present the whole time. In that mode, however strong the AI, it's essentially a tool waiting for orders — if you don't ask, it doesn't move.

What an agent (an AI system that executes multi-step tasks autonomously) breaks is precisely that "waiting". It doesn't stop after one reply. Given a goal, it breaks it into steps itself: look things up, write code, run it, hit an error, go back and fix it, then continue. The whole process may run dozens of rounds with no human touching it. That's no longer "responding to a prompt"; it's "executing a project".

Where do you see the difference? In whether it decides "what to do next" by itself. When a human uses AI, you decide every step; when an agent uses AI, it decides every step, and the human only sets the goal at the start and checks the result at the end. The middle — research, coding, browsing, debugging — is the AI working against its own output, not against a human prompt. It has gone from "the thing being used" to "the user that is using other AIs".

Why do agents burn so many tokens?

Ask ChatGPT a question and it computes one answer and stops: a one-off cost. Agents don't count that way.

Before doing anything, it first thinks "how should this be broken down" — one round of tokens. Then it decides which tool to call — search, run code, read a file — and the call itself costs tokens to describe what to do, and the tool's return costs tokens to read in. If it reads that and finds the direction is wrong, another round to re-plan. That hasn't completed a step yet; it has only completed "deciding how to do this step".

Once it actually starts, most tasks don't succeed first time. The code throws an error message and the agent has to read the whole thing to analyse the cause — that message alone may be hundreds of tokens. It writes a fix, the fix has to be re-verified, the verification result read again. A bug a human spots in three seconds may take an agent five to ten rounds of "run – read – analyse – fix" to converge.

A human question averages a few hundred tokens; an agent finishing one moderately complex task can stack up tens of thousands just in this internal self-talk loop. The 5× gap isn't because agents are doing five times as much. It's that to do the same thing, they have to talk to themselves five times as often to get it done.

One aside: the bill didn't grow 14×. Of that 7.3 trillion, roughly 70 to 85 percent are cache hits (the same context read repeatedly), which are billed at a discount. So "usage explosion" and "cost explosion" are two different things; keep them apart when reading the data.

If AI is its own biggest user, who defines where AI goes?

We used to tune a product by watching user behaviour: which button nobody clicks, which flow most people get stuck in, how long they stay. That data came from real fingers and real hesitation. But when four-fifths of the traffic is agents running by themselves, does that logic still hold?

An agent's "behaviour pattern" is nothing like a human's. It doesn't hesitate over the next step, doesn't bounce because the interface is ugly; it just runs the loop it was set until the task completes or the tokens run out. A team making product decisions on that data is using a machine's efficiency curve to guess at human habits — the two aren't even on the same coordinate system.

Worse is the question of who gets to speak. When the data is mostly generated by AI-to-AI interaction, who is entitled to say "this is the direction users need"? The engineers who wrote the agent? The person who launched the task? Or the AI itself, endlessly self-checking and self-correcting, that actually burned those 7.3 trillion tokens? We used to decide what AI should become by "observing users"; now the definition of "user" itself has come loose.

An uneasy question left for us

7.3 trillion tokens, five times human usage — this isn't a number that corrects itself. The curve runs one way, with no sign of slowing. The old chain was: humans use a product, the product collects human behaviour, the company adjusts direction from the data. The first link in that chain was always a person.

Now the first link is being replaced. When agents research, code and choose the next tool by themselves, what they produce is a record of "AI-to-AI" behaviour, not "human-to-AI" use. When we use that data to judge where AI should go, we're really watching machines serve machines, then deciding whether to keep serving people the same way.

If humans are no longer AI's biggest user,
what gives us the standing to say how AI should develop?

That question is worth ten seconds of thought before you next open ChatGPT or any agent tool.

Author：林政賢（Director · Gen AI creator & engineer · Founder of TangYi Studio）