A prompt has 200 input tokens marked as cacheable. Cache hits report no cost reduction. What is the most likely explanation?
Answer: The prompt is below the minimum cacheable size (typically ~1024 tokens) — caching has overhead and doesn't engage for very short prefixes
Anthropic's prompt caching has a minimum input size (typically around 1024 tokens). Caching a 200-token prompt provides no benefit because the prompt is below the threshold. The mitigation is to either cache only when the prefix is large enough, or restructure the prompt to push the cacheable region above the threshold.