What Is Prompt Caching and How Does It Work?

It lets a model recycle what it has already done when part of the prompt repeats. Those tokens cost you less and you get an earlier answer.

Prompt caching recycles the inputted prefix within one inference, thus making the inference less costly but still taking place. Semantic caching uses an already saved answer directly without making the model run whenever the new query is semantically the same.

No, because the machine produces an original output. Caching just ensures that the repeated prefix does not get re-evaluated; consequently, the quality remains unchanged, but the time and cost reduce.

Prompt caching relies on the prefix, which is the beginning of the prompt. Placing stable data at the beginning and changeable data at the end makes more of the prompt cacheable.

No, because the cache lives within a single provider. Different providers have different pricing and time-to-live policies. The caching is not carried forward when models or providers are switched.