What's the highest-leverage cost optimization for a production LLM app spending $10k+/month on tokens, BEFORE prompt-level token trimming?

The biggest cost lever for most LLM apps is not how many tokens you send, it is which model you send them to. Flagship models like Claude Opus 4.7 or GPT-5.5 are five to fifteen times more expensive per token than their smaller siblings like Claude Haiku or GPT-5-mini. Many tasks in a production app do not need a flagship. Simple classification, short extraction, routine generation, summarization of clean text. A small model handles these at the same quality you need. The routing pattern: classify each incoming request by complexity, send simple ones to the cheap tier, escalate to the flagship only when the task genuinely needs it. Teams who do this report cost reductions of fifty to eighty percent with acceptable quality. Prompt trimming usually buys ten to thirty percent at best, because the per token price stays the same. Start with the price tier, then come back to the prompt.

Cost optimization in a production LLM app has a clear order of operations: model routing first, prompt structure second, decoding params third. Token-by-token trimming is the last fifteen percent. The argument is multiplicative. A flagship like GPT-5.5 or Claude Opus 4.7 prices in at roughly five to fifteen times the per token cost of its smaller sibling. If you can route half your traffic to the cheap tier, you cut total spend by close to half before you touch a single word of the prompt. Compare that to prompt trimming, which works inside a single price tier and typically buys ten to thirty percent. The routing layer is itself a small engineering surface. A lightweight classifier or even a heuristic (length, format, presence of code) decides the tier. Hard cases escalate; easy cases stay cheap. Production stacks add a confidence check: if the cheap tier flags low confidence, the request gets retried on the flagship. The economics still win because escalation is rare. Once routing is in place, the prompt-trimming and cache-hit work become real and worth doing. But the first dollars come from the price tier, not the prompt length.