What is a token in AI?
A rate card priced per 1,000 tokens tells you nothing until you know how many tokens your documents actually are.
Token
tokenization
A token is a chunk of text a language model reads or writes as one unit, usually a word piece rather than a whole word. Anthropic and OpenAI both bill, limit and measure model behaviour in tokens, not in characters or words.
Before a model reads anything, a tokenizer breaks it into pieces from a fixed vocabulary. Common words often become one token. Long or unusual words split into two or three. Punctuation and spaces usually get their own token too.
OpenAI's own guidance puts one token at roughly four characters of English text, or about three quarters of a word. So 100 tokens is close to 75 words. Keep that ratio close by. Every price, limit and speed figure a model provider publishes is denominated in tokens, not words.
It is the first spreadsheet in any AI scoping
Before we discuss architecture on a call, we run the token math. Take a support inbox of 10,000-word documents, a policy manual of 40 pages, or a folder of contracts. Convert word count to a token estimate and multiply by the per-token price for input and output. That number tells you whether a project is a straightforward integration or a cost problem before a line of code is written.
It also explains latency. A model writes an answer one token at a time, so a longer reply takes measurably longer to stream back. A system prompt repeated on every call is tokens you pay for on every request, whether the user notices or not.
- 01A context window limit is a token count, so "can it read the whole document" is really a token question.
- 02Two models priced the same per token can still cost differently on your workload, because they split your language or code into tokens at different rates.
- 03A fixed cost, on every call: the system prompt.
- TextYour prompt and documents.
- TokenizeSplit into model-vocabulary pieces.
- ProcessCounted against the context limit.
- GenerateOutput tokens, one at a time.
- BillInput and output priced separately.
Input and output are almost always priced differently, and output usually costs more per token than input.
Related questions
01How many tokens is one word?
About 1.3 tokens per English word, or one token per roughly four characters, using OpenAI's published estimate. Short common words are often a single token; longer or rarer words split into two or more.
02Why do LLMs charge per token instead of per request?
Because compute cost scales with how much text a model reads and writes, not with how many requests you send. A one-line question and a fifty-page document cost very different amounts to process. Per-token pricing charges for the work actually done.
03Do input and output tokens cost the same?
No, usually not. Providers commonly price output tokens higher than input tokens, since generating text is more computationally expensive than reading it. Check the current rate card for the model you are using rather than assuming parity.
