Gemini context caching bills for storage duration even when the cache is never read

Explicit context caching creates a cachedContents resource holding a fixed prefix of tokens, which later requests reference by name instead of resending. Reads of cached tokens are discounted, but the cache is also billed for how long it is stored, per token per hour, for the whole time to live whether or not anything reads it.

That changes the economics compared to prompt caching on other providers, where an unused cache entry simply expires at no cost. Creating a cache per user session and never reading it twice can cost more than sending the tokens again.

There is also a minimum token count below which content cannot be cached at all, and it varies by model, so a prefix that seems large can still be refused. Set an explicit time to live matched to how long the prefix is actually reused, delete caches when the session ends, and measure before assuming caching is cheaper for your traffic shape.