Tokens

every bill, rate limit and context error you will ever see is denominated in tokens, not words or requests.

Token

the same string cost a different number of tokens on the other provider, because each model family ships its own.

Tokeniser

everything the model can ever emit is drawn from one fixed list of a few tens of thousands of entries.

Vocabulary

the model never sees your text at all, only the list of integers the tokeniser handed over.

Token ID

you notice a common word costs one token and a rare surname costs five, because frequent pairs got merged and yours was left in pieces.

Byte pair encoding (BPE)

the model cannot see individual characters unless the tokeniser happens to split them out.

Tokeniser quirks

the reserved markers for turn boundaries get inserted for you, and malformed ones are behind a good share of jailbreaks.

Special tokens

if a local model behaves bizarrely, a wrong or missing chat template is the first suspect.

Chat template

everything you send and everything it says has to fit one budget, and the retrieved documents are what usually blow it.

Context window

you halved the prompt and the bill barely moved, because the long generated answer was where the money actually went.

Input vs output tokens

the request fit fine and the answer still stopped mid-sentence, because the output cap is a separate and much smaller limit.

Context length vs max output

you count the tokens before sending rather than multiplying the word count by 1.3 and hoping.

Token counting

you time it yourself on your own hardware, because the number in the announcement was measured on something else.

Tokens per second