LLM Bills Have Up to Six Cost Lines — Most Developers Only Watch Two

AI language model pricing is typically advertised as two figures — input and output costs per million tokens — but real invoices can include up to five or six distinct billing lines. Hidden charges such as cached input, cache writes, reasoning tokens, and image or audio inputs often account for the bulk of unexpected costs. Reasoning-capable models silently bill for internal 'thinking' tokens at the output rate, even though users never see that content. Prompt size is identified as the most common driver of inflated bills, since every token in a request — including system instructions, tool schemas, and conversation history — is billed regardless of its usefulness. Accurate cost forecasting requires counting tokens with the specific tokenizer of the model being used, as tokenization rates vary significantly across model families and content types.