Prompt caching cuts repeated input-token costs up to 90%. How OpenAI, Anthropic, and Google differ, plus prompt …