The machine cannot see the word you typed.
Everything you are billed for, limited by and frustrated by comes down to one unglamorous step: your sentence is chopped into pieces before anything else happens.
The chopping comes first
Before a model does anything you would recognise as thinking, your sentence is cut into fragments. Not words — fragments. Common words usually survive whole. Rarer ones get broken into two or three pieces, and an unusual name might be split into five.
These fragments are the only thing the machine ever handles. It has no concept of a word, a sentence or a paragraph. It has a long row of pieces, and a very good statistical sense of which piece tends to follow which.
That sounds like a technicality. It is the reason behind almost every odd behaviour you have noticed.
Why your bill is shaped like this
Every provider charges per fragment, in and out. This is why the same question costs different amounts in different languages: English is the best-served case, and a language the tokeniser was less tuned for can take two or three times as many pieces to say exactly the same thing. Same meaning, same answer, materially different invoice.
It also explains why pasting a long document is expensive in a way that asking a long question is not. You are not paying for difficulty. You are paying for volume.
The window is a desk, not a library
A context window is quoted in these same fragments — a hundred thousand, a million. It is easy to hear that as a measure of how well-read the thing is. It is nothing of the sort. It is a ceiling on how much can be placed in front of it while one answer is produced.
Nothing inside it persists. End the conversation and the entire contents are discarded, so anything that must still be true next week has to be stored by you and handed back deliberately.
The useful picture is a desk. A bigger desk lets you spread out more papers while you work. It does not mean you have read more books, and it does not mean anything stays there after you walk away.
Where this bites in practice
Counting is a good example. Ask a model how many letters are in a word and it may well get it wrong, because it never saw the letters — it saw two or three fragments. The question is, from its position, a bit like being asked to count the bricks in a house you have only seen from the air.
What to do about it
Measure your real traffic in fragments rather than in words before you commit to a price. Check the ratio in every language you serve. And treat a large context window as more desk space, not as memory — if something must be remembered next week, write it down somewhere yourself.