Four multipliers chat users never see
Context is re-sent every step
LLM APIs are stateless. A 40-step agent session with 60k tokens of files, history, and instructions in context re-transmits that payload 40 times: 2.4M input tokens for one task — even though you only "chatted" once. Prompt caching discounts the cost, but the tokens are still processed and metered.
Tool output is input
Every file read, grep result, test log, and stack trace comes back into the model as input tokens. A single failing test suite can dump 20k tokens of logs into context — then get re-sent on all subsequent steps (multiplier 1 compounds it).
Retries and self-correction
Agents try, fail, read the error, and try again. A task that a human describes in 50 tokens can take an agent 15 attempts. Failure is cheap in dollars but expensive in tokens — and it's invisible in the "one prompt" mental model.
Parallel agents & background jobs
Power users run 3–10 agents concurrently, plus scheduled jobs (CI review, issue triage) that consume tokens while they sleep. Multiply everything above by the fleet size and 24h uptime.
Worked example: the honest napkin math
One refactoring task: 60k context × 40 steps = 2.4M input plus ~150k output. Eight tasks a day → ~20M tokens/day, or ~600M tokens/month for one engaged developer. Run three parallel agents and you're near 2B tokens/month — per person.
So when vendors report trillions of tokens processed, you don't need inflated-metrics theories to explain it: 100 developers at 2B/month = 200B tokens/month from a single mid-size customer. The skepticism in the original post is healthy — but the multiplication checks out. The right question isn't "are the numbers real?" — it's "how much of that burn is productive vs. retry waste?"