AI is using more and more AI. According to Peter Walker, an analyst at the model routing platform OpenRouter, February 6, 2026 may have been the last day humans consumed more tokens than AI agents.

Since then agentic token usage has grown 14x. Human usage over the same period is up just 2.8x. The gap is close to fivefold.

Why it grows so fast

The mechanism is simple and self-reinforcing. Agents increasingly work on their own over longer stretches, spinning up additional AI processes along the way. Where a human asks one question and gets one answer, an agent can make dozens of calls:

  • Planning — breaking the task into steps.
  • Tool calls — a separate request for each step.
  • Evaluation — checking whether the result is correct.
  • Retrying — repeating a failed step differently.

Token inflation actually started with reasoning models, which think longer before they respond — even when they shouldn't.

The number is not as expensive as it looks

Raw figures can mislead, and the analysis says so plainly. Nearly 70 percent of agent token usage comes from cached prompts, which are billed at much lower rates.

The reason lies in how agents work. Across a single task an agent resends the same context repeatedly — the system instruction, the tool definitions, the previous steps. Providers cache that repeated portion and price it at a discount. So 14x the tokens does not mean 14x the cost.

That makes the efficiency of agent frameworks a direct cost factor. The difference between a framework that finishes a task in thirty calls and one that takes a hundred lands on the user's bill unchanged.

Limits of the measurement

The platform the data comes from skews toward open-weight models, which tend to be less token-efficient than models from the major labs. The absolute numbers may therefore be specific to this platform.

But the direction of the trend probably holds everywhere. The rise in agentic usage is driven not by the platform but by a change in how software works.

Why it matters

This data changes how the industry's growth figures should be read. "Token usage rose by X" has until now been read as a demand indicator: more tokens, more people, more use.

That no longer holds. Much of the increase comes not from new users but from software becoming more talkative while doing the same job. An agent making a hundred calls to finish a task does not make that task a hundred times more valuable.

The distinction matters on the investment side. Compute demand forecasts rest on token volume, and how much of that volume is real work versus agent chatter is a separate question. As agent frameworks grow more efficient, the same work can be done with fewer tokens — meaning today's growth curve may reflect a temporary inefficiency rather than durable demand.