Prompt caching makes cached input tokens 10x cheaper than regular ones for OpenAI and Anthropic APIs by storing attention mechanism data like key-value tensors from repeated prompt prefixes.

This skips recomputation on subsequent requests, reducing time-to-first-token latency by up to 85% for long prompts, as shown in tests with GPT-5 and Sonnet 4.5. The process involves tokenization into integers, embedding, and transformer layers where caching occurs, enabling faster inference without reusing full responses.