The Multilingual Cache Tax: How token-count cache thresholds create a non-monotonic cost bias across languages in LLM serving

Commercial LLM APIs price prompt caching (the reuse of an already-processed prompt prefix) as a first-class billing dimension, and every major provider gates the discount behind a minimum token threshold: a prefix caches only once it exceeds a fixed number of tokens, commonly 1,024. Separately, the same content costs a different number of tokens in different languages (token inflation). This paper shows that the interaction of these two facts produces a non-monotonic cross-lingual cost bias. Near the threshold, a higher-inflation language crosses into cache-eligibility at a shallower depth of content than English, so a service in Korean or Vietnamese can be several times cheaper than the same service in English at an identical content depth. Once prefixes are long enough that every language caches, the language-neutral discount preserves the inflation ratio, and high-inflation languages remain roughly 1.3 to 1.8 times more expensive. We give a provider-agnostic cost model, measure token inflation on FLORES-200 across four tokenizer families, characterize the crossover band where the bias flips sign, and confirm the 1,024-token floor, 128-token block alignment, and the English-caches-last paradox against live OpenAI billing. Code and data are included for reproducibility.