Monthly leaderboards now show autonomous agents burning tens of trillions of tokens. Each orb below is an app category sized by monthly token burn — tap orbs to inspect, and use the sliders to translate trillion-token scale into dollars, books, and human lifetimes.
A chat reply might use 2k tokens. One agent step re-sends the entire conversation + tool outputs as input every iteration. A 50-step coding session can consume 5–10M tokens — 2,500× a chat message. Persistent-memory agents with 40+ tools multiply this further.
Agent workloads are typically 90–98% input tokens (context re-reads) and only 2–10% output. That's why providers price input 3–5× cheaper and why prompt caching — reusing identical context prefixes — can cut agent bills by 50–90%.
Skeptics note token counts can be gamed: retries, cache reads counted as usage, and synthetic-data generation all inflate figures. Healthy skepticism: check whether a leaderboard counts cached input separately before treating tokens as adoption proof.
Books assume ~90k tokens each (≈70k words). Human reading: 250 words/min ≈ 333 tokens/min, no sleep. Cost = tokens × blended $/1M. At 26T tokens and $0.80/1M, that's $20.8M/month of compute.