Guide
How to reduce LLM API costs
Reducing LLM API cost starts with measuring input tokens, cache reads, cache writes, output tokens, and request frequency. Context compression is one lever, not a universal answer.
Find repeated input first
Stable prefixes may benefit more from provider caching than from changing the prompt. Inspect cache behavior before trimming them.
Compress changing context
Verbose tool output and retrieved data are often better candidates for extractive selection before a model call.
Calculate net savings
Subtract Tokz credits and account for cached-input prices; label estimates and unknown provider data explicitly.