Savings

exactly what the caller declared for this scope.

hits served in this scope since the cache was created. Every one of them is a model call that was not made; a hit is the only event that adds to a saving, so a cache that is filling up reports nothing until it starts answering.

prompt tokens read from the served entries' metadata, summed.

completion tokens read from the served entries' metadata, summed.

hits whose entry carried neither token count. A number close to hits means the metadata keys are not the ones being written, and amount is then only the flat per-call charge rather than the saving. That is the one way this figure can be quietly wrong, so it is reported next to it rather than left to be discovered.