Cut your LLM bill by 30 to 70%: the levers that work

On the bills I audit, the problem is almost never the price per token. It is useless context sent on every call and the most expensive model plugged in everywhere by default. Here is what I cut first. In nearly every engagement, the problem is not the price per token but the way the tokens are spent: useless context sent back on every call, the most expensive model used everywhere by default, answers regenerated when they already existed. So I start by measuring where each euro goes, not by cutting at random. Yes, and it is almost always the first lever: in production, a large share of calls are near-duplicates. Caching the answers on identical inputs removes that waste without changing anything for your users, often within a few days. No. Routing each request to the cheapest model capable of handling it is enough in most cases: a simple classification or extraction does not need the most powerful model. I never cut a bill by degrading quality. I cut it by no longer paying for what adds nothing. Cut the dead context: everything the model never reads still costs money. Favor short instructions and targeted examples over long directives. Group bulk processing into batches when latency allows. Taken together, these levers bring a bill down by 30 to 70% on most of the products I audit, at equal quality, and the first gains usually land within two to three weeks. How fast does the LLM bill drop? Does cutting the bill mean degrading quality? Do I need to switch model provider to save? I write about shipping AI to production at guinat.ai. Honest advice, no hype.