I measured this on my past claude code sessions. I have a dataset with full LLM requests and responses and I ended up to an estimated 5-7% savings on tokens. Al

I measured this on my past claude code sessions. I have a dataset with full LLM requests and responses and I ended up to an estimated 5-7% savings on tokens. Also with my primarily Opus 4.8 usage, only 5.9% of unique reads even qualify for the bulk-read operation and 58% of the reads were already targeted.