Five ways to cut prompt cost without losing quality
Prompt cost is one of the few line items you can reduce by editing text. The catch is that aggressive trimming degrades answers, so the useful question is which cuts are free. These five are.
1. Delete the politeness scaffolding#
"I was just wondering if you could please help me to" is eleven tokens that carry no instruction. Models do not respond better to courtesy, and phrases like in order to, due to the fact that, and at this point in time each collapse to a shorter equivalent with no change in meaning. The Quick clean-up button applies exactly these substitutions.
2. Cap the output, not just the input#
Output tokens cost several times more than input tokens, so "answer in under 100 words" frequently saves more than rewriting the entire prompt. If you need structured data, ask for the structure directly rather than prose that happens to contain it.
3. Stop resending context that hasn't changed#
In multi-turn applications, the full history is re-sent on every request. Ten turns of a conversation with a large system prompt means paying for that prompt ten times. Prompt caching exists for this and typically cuts the repeated portion to a fraction of full price — but only if the cached prefix is byte-identical, so a timestamp interpolated into your system prompt quietly defeats it.
4. Send the excerpt, not the document#
Attaching a whole file when three paragraphs are relevant is the most common source of waste. Retrieve the relevant span first and send that. This usually improves answers too: models attend less reliably to material buried in a long context than to the same material presented on its own.
5. Match the model to the task#
Classification, extraction, and routing rarely need a frontier model. The price difference between tiers is often an order of magnitude, which dwarfs anything you will achieve through prompt editing. Compare the per-model cost for your own text in the table before optimizing the wording.
Measure before and after#
Every one of these is easy to over-apply. Paste your prompt into the calculator, run the clean-up or the AI rewrite, and check the savings figure against the model you actually use. If a rewrite saves four percent, it is not worth the risk of changing behavior — spend the effort on points three and five instead.