Comment by naaqq

DeepSeek’s official API has a cache hit rate of over 99% if you use it continuously within the same codebase for long sessions, so it’s much cheaper than frontier models. I have an example of 200M token session in claude code.