We pay for AI by the word.
Every time you use AI, you pay for what you send in and what it sends back. More words in and out means more cost.
When you use AI as an assistant, you ask, it answers, and you make the next call. That's why a copilot-style user genuinely struggles to spend more than $10 to $15 a month. They're doing the thinking, the AI's helping.
Agents are a different animal. You hand over a goal ("research 20 competitors, compare pricing, write the report, check the sources") and the AI quietly runs dozens or hundreds of steps to get there. Read, think, search, write, check, retry. Every one of those steps is more words, and more words is more money.
One agent run can sit around $3.50 instead of a few cents.
A skilled operator gives the agent clear scope, limits and checkpoints. An unskilled one says "just keep going until it's good" and walks away. That's how you go from a sensible daily spend to genuinely astronomical bills.
The cost isn't really about the tool. It's about how much rope you give it.
A skilled AI user spends what they need in tokens, not what they can.
