Bad prompts cost more than tokens
In a notebook, a bad prompt costs a retry. In production, it costs:
- User trust — a confidently wrong answer once is worse than a refusal ten times.
- Tool budget — vague instructions cause loop-y tool use, multiplying API and latency cost.
- Reasoning budget — "think step by step" on a trivial task wastes thinking tokens, which are billed.
- Downstream code — malformed JSON breaks parsers; you pay in incident pages, not credits.
- Compliance and brand — one prompt-injected response that leaks PII, contains slurs, or makes a legal claim is the most expensive token you'll ever generate.
The hierarchy of failure cost
Prioritize prompts in proportion to their blast radius. A prompt that drives marketing copy generation is not the same as one that approves a refund. The same words have different weight depending on what they touch downstream.
Where the savings hide
The biggest cost lever in 2026 is not the model choice; it's how much context you pack and how much reasoning you ask for. A clean prompt with tight context can cost an order of magnitude less than a sloppy one — and produce better outputs because the model isn't drowning in noise.
GPT -5.5 기준으로 보면 input 5달러/1m 토큰, output은 30달러 /1m토큰이다. 비용을 줄이는 가장 큰 효과는 output을 줄이는것이에요. 그런데 아웃풋을 줄이면서 동시에 퀄리티를 망치지 않으려면 아웃풋을 줄이는 것인데, 미사여구들, 설명하는 쿠션언어들을 제외시키는 것으로 생각했어요.
그런데 그러면 여기서 발생하는 의문은, 결국 아웃풋의 형태는 structured output만 남는거 아닌가 싶어요. 사실 아웃풋에는 쿠션언어들이나 설명문도 필요한 거 아닌가요?