Still on 2.1.87, exclusively Opus for coding — haven't hit this yet. Wondering if the bug is personal vs team plan specific?
I'm sure it's more complex, but why not improve internal implicit caching and pass the savings on? Presumably Anthropic already benefits from caching repeated prompt prefixes internally — just do that better, extend the TTL window, and let users benefit. Explicit caching stays for production use cases with semi-static prompts where you want control.
The current 5-min default TTL + 2x penalty for 1-hour cache feels punitive for an interactive coding tool.