A context window is a limit, not a budget.

That distinction gets lost once windows reach hundreds of thousands of tokens. The cost of a request follows what you actually send, not how much room was available. A bigger window doesn't make an abandoned branch free. It just lets you carry three of them before anything visibly breaks.

In a long working session, most of what piles up isn't the active task. It's decisions that were later overturned, approaches you tried and dropped, the same instruction restated because it wasn't clear the earlier one still applied, and a schema pulled in for one question that is still loaded twenty exchanges later.

All of it gets processed alongside the material that matters. That shows up in the bill, and it can show up in latency and output quality too.

A quick test: if you started this exact task fresh, how much of what's loaded right now would you bring with you?

I wrote up the longer breakdown here: https://medium.com/@nagatomopedro05/the-bigger-context-window-didnt-save-you-any-money-here-s-the-receipt-73f6236ee607