Cost and tokens
Most cost advice confuses three separate things: fewer tokens, a cheaper model, and fewer attempts. They are not interchangeable.
What you are charged for
Providers bill per token, separately for what goes in and what comes out, with output typically several times more expensive than input. A token is roughly three-quarters of an English word, but the ratio differs by language and gets worse for languages that are not English. Structured data, code and unusual names all cost more tokens than plain prose.
Two things follow. First, a long system prompt sent on every call is a recurring cost, not a one-off. Second, asking for a shorter answer often saves more than trimming the input.
“Caching makes the prompt shorter.”
What is actually trueCaching makes repeated input cheaper and often faster. Those tokens are still part of the model's context and still shape the answer.
“The cheapest model gives the lowest cost.”
What is actually trueA cheap model that needs three attempts, extra tool calls and human cleanup can cost more per accepted result than an expensive one that works first time.
“Reasoning tokens are free because you cannot see them.”
What is actually trueHidden thinking is billed as output on every major platform. A high effort setting on a simple task is a real and invisible line on the bill.
Where the savings actually are
Cap the output
Output is the expensive side. Asking for a 200-word memo instead of an unbounded one is usually the single biggest saving available.
Match effort to difficulty
Current models expose a reasoning effort control. Routine extraction does not need maximum effort; use the dial rather than the prompt.
Cache the stable part
Keep the unchanging instructions and reference material at the front of the prompt so it can be cached, and put the variable part last.
Reduce attempts, not just tokens
A brief that works first time beats a shorter one you have to run three times. Measure cost per accepted result, not cost per call.
Route by task
Use a small model for classification and extraction, and reserve the frontier model for work that genuinely needs it.