Most of us picture an AI service as a chat window. You type a question, it answers, and whether you’re on a free account or paying a monthly fee, you rarely see a ceiling. So it can be puzzling to hear that someone managed to exhaust a $150 monthly plan before the middle of the month. To understand how that happens, it helps to know what these tools are counting in the first place, because it isn’t the number of times you hit send.
The unit is the token
AI services don’t tally your questions one by one. They measure tokens, and a token takes some getting used to.
A token is a small piece of text: the unit a language model breaks your writing into before it does anything with it. The awkward part is that it has no fixed size. A page is roughly 500 words no matter what’s printed on it, but a token isn’t that dependable. It might be a whole word, part of a word, or a single mark of punctuation. Ordinary words often become one token; running is one. Longer or less common words get split into pieces, so tokenization comes out as something like token and ization, and the apostrophe in don't can count on its own.
The model isn’t reading for meaning while it does this. A tokenizer simply cuts the text into pieces of a size the system was built to handle, treating everything the same way: words, spaces, brackets, punctuation, even portions of an image if you hand it a picture. As a loose guide, 1,000 tokens works out to about 750 words of ordinary prose. That ratio drifts, though, and anyone who works with digitized collections will see why immediately. Structured data, tables, older orthography, non-English text, and anything with unusual characters all shift the count. A page of Fraktur or a rough OCR transcription behaves nothing like clean modern English, and the token count reflects that.
There’s one detail that catches almost everyone off guard. The model doesn’t only read your latest message. Every time you send something, it re-reads the whole conversation from the beginning. A long back-and-forth isn’t a string of small, separate tasks; it’s an accumulating transcript that gets processed again, in full, on every turn.
From tokens to credits, money, and energy
Different products dress this up differently. Some show you credits: Perplexity, for instance, gives a daily allowance for heavier features like deep research or web search. Others show a percentage bar, or just a quiet note that you’re running low. The label on the dashboard changes from tool to tool, but the question underneath is always the same. How many tokens did this take, and what did each one cost to process?
Money follows from that count, and so does energy. Tokens are a rough proxy for electricity. A short exchange uses a tiny fraction of a watt-hour. A long session grinding through a large document can use more than a month of casual questions put together. For anyone who thinks about the environmental cost of this technology, and it’s a live question in our field, the useful measure isn’t how often you opened the app. It’s how much work each request actually demanded.
Where the numbers climb
Most people never come close to a limit. The ones who do usually aren’t doing anything exotic. They’ve simply moved into a heavier kind of use, and the clearest example is close to home.
Picture someone running transcription across a digitized collection rather than asking the occasional question. Instead of a sentence or two, they’re handing the model whole documents at a time: scanned pages, existing transcriptions, metadata records, a page of instructions about how to handle it all. A single page might be a few hundred tokens. A batch of dense newspaper pages, or a run through hundreds of items in a finding aid, is a different order of magnitude. The model reads all of it, answers, and then reads all of it again on the next request. Across a working day of processing, correcting, and reprocessing, that reaches into the millions of tokens without any drama at all.
This is why a $150 plan can still run dry. It buys a generous but fixed budget, and someone working through a large, complicated project (long documents, multi-step reasoning, repeated passes to fix errors) can spend it in a week or two. Same figure on the invoice as a casual user, an entirely different relationship to the limit.
There’s a further wrinkle with the newer “reasoning” models, the ones that work through a problem in steps before answering. That extra deliberation genuinely helps on hard tasks, and it genuinely costs more, because each intermediate step is itself text being generated and then read back. The more thinking you ask for, the more you spend.
It’s also worth knowing that these budgets don’t combine across tools. Each service meters only its own work. If you use one tool to build or run a workflow, and that workflow in turn calls out to a second AI service to read images or pull text off a page, you now have two meters running at once. One counts the work of the tool in front of you; the other counts whatever your process quietly sends it. Someone running an automated pipeline isn’t so much a heavy user of a single service as a person drawing on several budgets at the same time, some shown plainly on a dashboard and some tucked away inside the process they’ve set up.
For everyone else
The free and entry-level tiers cover what most people actually do: short exchanges, an occasional document, a question when one comes up. They’re set high enough that the ceiling stays out of sight long enough for the tool to earn its place. By the time you bump into a limit, you’ve usually already worked out what you want it for. At that point the meter isn’t really a shock. It’s a sign the thing has become useful.