What a token measures
A language model does not read letters or whole words. Before anything happens, your text is split into tokens: common words usually become a single token, rarer or longer ones get broken into pieces, and punctuation and spacing count too. The model then works entirely in those pieces, predicting the next one over and over until the answer is finished.
Two consequences follow, and both cost money. Every request is charged by how many tokens go in and how many come out, so a long document pasted in and a long answer requested are both expensive. And every model has a ceiling on how many tokens it can hold in one conversation — its context window. Reach that ceiling and something has to be dropped, which is usually the earliest part of the conversation.
Why tokens matter
They turn vague worries about AI cost into something you can plan. If you know roughly how much text goes into a task and how much comes back, you can estimate what running it across a catalogue or an inbox will cost before you commit to it, rather than discovering it on an invoice.
They also explain behaviour that otherwise looks like a fault. An assistant that seems to forget an instruction given earlier in a long chat has not malfunctioned; the instruction has fallen outside what it can still see. An answer that stops mid-sentence has hit an output limit. A tool that gets slower and dearer as the conversation grows is doing exactly what it should, because every reply re-reads everything before it.
There is a point that matters in Nepal and anywhere multilingual. Languages written in scripts less represented in a model’s training tend to be split into more tokens per word than English. The same message in Nepali can cost noticeably more to process than its English translation, and it eats the context window faster.
Common mistakes with tokens
Confusing tokens with words when budgeting is the routine one, and it always errs in the wrong direction: the token count is higher than the word count, more so for names, product codes, URLs and non-English text. The second is pasting whole documents in when a relevant extract would do, then paying for the same background material on every follow-up question.
The third is letting one long-running conversation carry everything. Once the earliest part starts falling out of view, the model’s answers change without warning, and a chat that ran for hours quietly stops following the rules it was given at the start.
How to act on it
Send only what the task needs. Extract the relevant section rather than attaching the whole file, ask for the length of answer you actually want, and start a fresh conversation for a new job instead of continuing an old one out of convenience.
For anything you will run repeatedly, measure one instance first — check the usage figures in your provider’s dashboard — and multiply by realistic volume before rolling it out. Where a cheaper, smaller model does the job, use it for the bulk work and reserve the expensive one for the parts that need judgement. That kind of sizing is ordinary discipline in any automated workflow.