How to cut your OpenAI bill by 60% without touching your prompts
No prompt engineering, no model downgrade, no accuracy trade. Just stop sending the model questions it has already answered in slightly different words.
Most cost-cutting advice for LLM apps asks you to give something up, a smaller model, terser prompts, fewer tokens, worse answers. There's a fatter, lazier win hiding in plain sight that costs you nothing: a huge share of your traffic is the same questions, reworded. You're paying full freight to answer them again and again.
The same answer, billed four times, until the cache understands the wording is irrelevant.
On a workload that's roughly two-thirds repetitive, support bots, FAQs, agents re-asking for the same schema, that's a 60-70% cut, and it arrives without a single change to how you write prompts. It's the rare optimization that's also less work: you delete calls instead of engineering them.
The bottom line
The cheapest token is the one you never send. Stop paying twice, keep your prompts exactly as they are, and let the cache quietly eat the duplicates.