A token is the atomic unit of text that a language model processes. One token typically corresponds to about four characters, though the exact amount varies. section 1.5.1
Models read tokens from inputs and generate tokens in outputs. Processing fees are often tied to token counts, so counting tokens helps estimate invocation costs and compare model choices. section 1.5.1 section 9.2
Estimating review-classification costs
Suppose a product review averages 100 tokens, or about 400 characters, and a company receives 10,000 reviews daily. That amounts to roughly one million tokens per day. At the example prices given, processing a year of reviews costs about $21,900 with GPT-4 or $182.50 with GPT-3.5 Turbo. section 9.1
