12:31
2026-09-24
pub.towardsai.net
large-language-models
GPT-5.6 Quietly Broke Our Prompt Cache. One Message Boundary Fixed It.
A model upgrade to GPT-5.6 (Sol, Terra, Luna) dropped one team's prompt cache hit rate from roughly 91% to near zero because the new caching model only stops at the end of an eligible message, not at …