You ask your AI assistant to do something that feels like the same kind of request you’ve always made, and this time it blows through your monthly limit before lunch. That’s not your imagination. It’s not that your provider quietly shrank your quota. The request itself is doing something completely different under the hood, and almost none of that work is visible to you.
I run my own system prompt in Claude that tells it not to write anything into a Microsoft file unless I specifically ask for it. I added that instruction because I kept noticing the assistant defaulting to habits I hadn’t asked for. That small rule is a good entry point into the bigger idea here: everything you tell a model, plus everything it writes back, gets fed back in on every single step. Nothing is remembered the way a person remembers a conversation. It’s re-read in full, every time.
How AI actually writes
Here’s the simplified version. Ask a question like “What did the cat sit on?” and the model doesn’t produce the full answer in one motion. It writes one token — say, “The” — then feeds the entire question plus that one word back through itself to write the next token, “cat.” Then it feeds the question plus “The cat” back through to write “sat.” Every single word costs a full pass over everything that came before it. For a seven-word sentence, that’s seven separate passes, each one re-reading more than the last.
That mechanic barely matters when you’re asking a simple question in a chat window. Load the system prompt, load your question, get a response — around 1,300 tokens total, once. That’s the world most of us built our intuition for AI cost around, and it’s why AI felt cheap for the first year or two people used it seriously.
Then you hand it a real task
Ask it to build a Q3 proposal deck using your standard template, the discovery call transcript, and your meeting notes, and the picture changes completely. Now it isn’t just answering — it’s opening files and reading them in full, drafting new slides, rendering those slides to images so it can visually check its own formatting (it genuinely can’t judge layout reliably from raw XML alone), spotting a problem, editing, and rendering again to confirm the fix worked. None of that comes back to you. You only see the final “here’s your updated deck.”
By the time it’s done, that single request has burned through roughly 163,000 tokens — for what started as one message.
Why this scales the way it does
The part worth sitting with is that this doesn’t scale the way you’d expect. It isn’t linear with how much you asked for — it’s exponential with how many things the model decides it needs to do. Connect this same pattern to your actual systems — a tool call into your file structure, a query against your ERP — and every one of those steps reads its own results back in, then potentially triggers another step. A request that looks simple on your end can spiral into dozens of hidden passes before you ever see an answer.
This is likely why your token usage has jumped even though you feel like you’re asking for the same kind of help you always have. You’ve probably moved from simple chat-style questions to agentic requests — attaching files, asking for multi-step deliverables — without the cost model in your head keeping pace.
Key Takeaways
- Every forward pass re-reads the entire context, not just the newest input — nothing is remembered for free.
- A plain chat question stays cheap because nothing gets fetched or re-checked; an agentic request with files and tool calls compounds with every step.
- Cost scales with the number of things the model decides to do, not the size of your original ask — a “simple” deck request can spend well over 100x what a single paragraph costs.
- Rendering-and-reviewing loops, like checking slide formatting by generating images, are invisible to you but billed like everything else.
- Before connecting AI to your file systems or ERP, expect each additional tool call to multiply, not just add to, your token spend.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply