Comment by lifthrasiir

You are right to be curious. The encoding used by both GPT-3.5 and GPT-4 is called `cl100k_base`, which immediately and correctly suggests that there are about 100K tokens.