Price the workload before the invoice does.

• Cost per request is input tokens times the input price plus output tokens times the output price, both per million tokens. Monthly volume assumes a 30-day month.
• Optimisations stack in a fixed order: semantic cache first, then model routing on the cache misses, then prompt compression and the output cap on whatever still reaches a model.
• No double counting. Each saving is measured against the cost left by the steps before it, so the parts always add up to the total. That is also why four optimisations together save less than the sum of each one on its own.
• Money is rounded to the cent only at the end; peak requests and tokens per minute are rounded up, because a partial request still counts against a limit.

• Prices change. The defaults are illustrative list prices for Anthropic models at the date shown. Check the provider, and include any discounts, batch pricing or prompt caching you already have.
• Rate limits and latencies are illustrative round numbers, not any provider's real values. Real limits depend on your account and usage tier, and some providers count input and output tokens separately.
• Cache hits are treated as free. A real semantic cache costs something to run: embedding each question, a vector store and the gateway itself.
• The output cap trims the average answer. A real cap only cuts long answers, and cutting them can hurt quality, so measure before you set one.
• Routing only works when the cheaper model is good enough for the requests you send it. Measure answer quality per route, not just cost.

Caching, routing, compression and token limits are all policies that sit between your applications and the model providers. Putting them in one gateway means every team gets them without rewriting their code, and the platform team can see cost and headroom for all traffic in one place.