GPT-5.6 prompt caching reportedly broken

Users report that GPT-5.6 prompt caching charges the 1.25× cache-write fee but never credits cached tokens, even with plain API calls. A Reddit user demonstrates reproducibility with two curl requests.