PSA - Claude Code: Turn off Prompt Suggestions, save ~10% of your limits/spend

tl;dr Prompt Suggestions - the greyed out "helpful" text at the end of a prompt in your input line - do a Cache Read of your entire context to generate something like "commit and push", or "Start with 1A" . This is expensive - can be up to 10% of your weekly Fable limit. Short story - I was instrumenting my claude code to see what our monitoring tooling could find based on instrumenting claude code, ahead of a customer call. I found it useful enough to share here. (another would be - don't compact after the cache expires). Details in screenshots, it's pretty obscene when you get to very high context length. Prompt Suggestions can be as expensive as your actual prompts. Screenshot descriptions: 1. Main finding 2. Confirming that it isn't a "cheap model" that does this. 3. What type of tokens are used? 4. A compaction oopsie (and the cost absurdity visible) 5. List of normal turns (prompts) + their related suggestions & costs. Slightly biased to the expensive ones. Median: suggestions cost me 91% of normal prompt cost ! (Note - this is only the normal turns - not tool calls that the model invokes or writing files or doing other things to build your apps. You're not getting double your usage by turning this off, sorry ) 6. List of suggestions claude made. 69x "keep going" (it copied that from my earlier prompts)