With OpenAI at least you can specify the cache key and they even have this in the docs:

Use the prompt_cache_key parameter consistently across requests that share common prefixes. Select a granularity that keeps each unique prefix-prompt_cache_key combination below 15 requests per minute to avoid cache overflow.