GPT-5.6 prompt caching appears completely broken: you pay the new 1.25× cache-write fee, but cached_tokens never credits — reproducible with two curls
GPT-5.6 prompt caching appears completely broken: you pay the new 1.25× cache-write fee, but cached_tokens never credits — reproducible with two curls — reported by reddit.com, aggregated and ranked by ClawDigest.