How LLM Tokens Work (And Why They Explain Your AI Bill)

Short summary

Large language models process text as tokens (roughly three-quarters of a word each), not words—this design choice explains both their capabilities and your surprise AI bills. You pay per token in both directions: input includes your prompt, conversation history, and tools; output is priced higher. Understanding tokenization is critical for anyone building or budgeting for LLM applications.

• •Tokens are fixed-dictionary chunks (words, subwords, punctuation), not individual characters or words
• •APIs price both input tokens (cheaper) and output tokens (3–5× more expensive), so costs compound with context size and history
• •Cost = (input tokens × input rate) + (output tokens × output rate); managing token count is the key to controlling spending

Generated with AI, which can make mistakes.