Cost planning guide
Input Tokens vs Output Tokens
AI bills usually separate what you send to the model from what the model writes back. That split matters because output tokens often cost more than input tokens.
Input tokens
Input tokens include prompts, system instructions, chat history, retrieved RAG chunks, uploaded files converted to text, tool results, and examples included in the request. A short user question can become expensive when it carries a long document, a large codebase, or a full conversation history behind it.
For example, a customer support bot may receive a 20 word question, but the RAG system might add 10 document passages, policy instructions, output format rules, and recent chat history. The visible message is small, while the paid input can be thousands of tokens.
Output tokens
Output tokens are the answer generated by the model: text, code, JSON, tables, report sections, or rewritten content. They matter because many providers price generated tokens higher than input tokens. Long explanations, repeated boilerplate, verbose Markdown tables, and multi-step reports can become the main cost driver.
Simple calculation pattern
A practical planning formula is: input tokens plus expected output tokens, multiplied by workflow rounds, multiplied by daily or monthly usage. Output can be estimated as a ratio of input. A short summary might be 10-15% of the input, a detailed analysis might be 25-50%, and full report generation can approach or exceed the input length.
Example
A workflow sends 80,000 input tokens and generates a 20,000 token answer. One run uses about 100,000 total tokens. If the workflow has 3 rounds and runs 20 times per day, monthly volume is roughly 180 million tokens before provider-specific tokenization differences.
What to control first
Reduce repeated input before changing models. Summarize old chat history, retrieve fewer but better document chunks, remove unused file types, and keep system instructions concise. Then control output with clear answer formats such as "return five bullets" or "write a 300 word summary."