This reframing really landed for me — especially the point that the cost that matters is cost per correct answer, not cost per million tokens. The retry example stuck with me: a cheap model that needs three attempts can quietly end up more expensive than a pricier one that gets it right first time. I also liked the practical order you suggest, fixing prompts and context first before reaching for a bigger model, since that is where a lot of wasted calls seem to come from. Do you find teams usually discover this from a bill shock, or from actually instrumenting latency and retry rates?