Every major LLM API now caches prompt prefixes: send the same beginning twice and the second call is cheaper and faster. What vendors rarely say is how long a cached prefix survives. Anthropic documents it (5 min, or 1 h on request); OpenAI gives a range ("5 to 10 minutes of inactivity, up to an hour", 24 h on opt-in); DeepSeek says "hours to days". Z.ai, Moonshot (Kimi) and Mistral publish nothing. I could not find a page that measures it.
So I measured it.