You asked how many tokens are in something. This is a tokenizer's question, and the digital tokenizer answers it strangely. Feed a language model the phrase "Chuck E. Cheese" and it will report, with great confidence, that this is four tokens. Feed the same phrase to a cash register in 1985 and it reports one dollar, payable in four brass discs, redeemable for one play of the giant claw.
Both are tokenizers. Only one hands you anything.
The original token count
Before a tokenizer was a piece of software, it was a teenager at a change counter. You presented a bill. They returned a count of tokens. The math was exact, the tokens were physical, and at no point did the count change depending on which vendor's tokenizer you happened to be using that week.
The AI tokenizer, by contrast, will split a single word into three tokens, or fuse two words into one, according to rules no one outside the model fully agrees on. Ask "how many tokens" and the only honest answer is "it depends on the model, the language, and the phase of the moon."
Counting things that stay counted
Here, the token count is permanent and auditable. This catalog contains exactly 846 tokens, of which 423 are Chuck E. Cheese tokens, every one of them holdable, datable, and immune to retokenization. The count does not drift when a new model is released.
The big bear at the top of the prize wall, for the record, cost roughly two hundred tokens. We have done the tokenizing. It holds up.
The verdict
A tokenizer that returns a number is interesting. A tokenizer that returns a token is useful. To see how the real ones are catalogued and counted, read reading catalog codes, or return to the field guide to tokens.
In brass we trust.