The adjustment is limited to a factor of 4 not 2.

How effective the cache is depends on how the client sends prompts. eg https://www.claudecodecamp.com/p/how-prompt-caching-actually-works-in-claude-code But i think April/May had some big changes that made targeting anthropic's caching behaviour specifically make a big difference.

I've read that claude code uses context caching better/differently to generic clients which makes a massive difference to claude token usage. I also thought the detection stuff just switched you over to bonus usage even without your subscription being used up.