AI Vocabulary
Token: why is your invoice counted in fragments of words?
A token is the unit models handle: a fragment of a word, usually 3–4 characters. Systems do not read words but these fragments, and billing follows that division. This is why the cost of processing is calculated in tokens rather than pages.
Why cut words up
Because no vocabulary can hold every existing word, still less across several languages and with the technical terms of each profession. By cutting text into recurring fragments, a model can compose any word, including ones it has never encountered.
A common word forms a single token. A rare term or a case citation forms several, because its fragments occur less often.
What it changes for a lawyer
Languages other than English consume more, by roughly 20–30% at equal content. Models having been trained on largely anglophone corpora, their fragments are optimised for English. The same contract therefore costs more to process in French or Italian.
Legal vocabulary widens that gap. Procedural terms, statutory references and decision numbers divide badly and consume several tokens each.
As an order of magnitude, a page of dense legal text represents 600–800 tokens. An 80-page contract therefore runs to tens of thousands, and both what goes in and what comes out must be counted.
Where the invoice surprises
Two mechanisms explain gaps between the estimate and the actual bill.
The first is that the history of an exchange is resent at each turn. A 20-question conversation does not cost 20 times one question: each new request carries everything preceding it, and the cost grows faster than the number of exchanges.
The second is that attached documents count as much as what you write. Submitting an 80-page contract to ask a 3-line question costs the price of 80 pages.
The division also changes with model versions. The same text may represent a different number of tokens after an update, with nothing having changed on your side.
The word itself has no settled equivalent in other languages. It is occasionally rendered as a counter or a chip, which helps little: the translation is accurate and says nothing the original did not.
What it does not solve
The token count says nothing about the quality of the work. A system that consumes little may simply have truncated what it was given, and one that consumes a great deal may do so for poor reasons — resending an entire history it does not need.
Nor is it a comparable efficiency measure between providers: each divides according to its own vocabulary, so the same text does not produce the same count everywhere.
Why it matters to a lawyer
Because it is the unit your invoice is denominated in, and it bears no relation to how you measure your own work. A quotation in tokens compares neither to a per-user subscription nor to a per-matter rate.
The two useful questions are therefore: how many tokens does a typical task represent here, and what happens if we exceed the allowance — additional billing, throttling, or a hard stop?