prompt caching 到底能省多少钱?机制、验证脚本和六种让缓存失效的写法(2026)

{
  "model": "你的模型名",
  "max_tokens": 256,
  "system": [
    {
      "type": "text",
      "text": "这里是很长的系统提示词……(要超过最小长度门槛)",
      "cache_control": { "type": "ephemeral" }
    }
  ],
  "messages": [{ "role": "user", "content": "这里是每次都不同的短问题" }]
}

import json, urllib.request

BASE  = "https://cocodot.co/api/ai"   # Anthropic 原生,不带 /v1
KEY   = "你的-key"
MODEL = "你的模型名"

# 造一段足够长的稳定前缀,确保超过最小可缓存长度门槛
PREFIX = ("你是一个严谨的代码审查助手。以下是团队的编码规范,请严格遵守。\n"
          "规范条目:每个函数必须有类型注解;禁止裸 except;日志必须结构化。\n") * 220

def call(question):
    body = {
        "model": MODEL,
        "max_tokens": 64,
        "system": [{"type": "text", "text": PREFIX,
                    "cache_control": {"type": "ephemeral"}}],
        "messages": [{"role": "user", "content": question}],
    }
    req = urllib.request.Request(
        BASE + "/v1/messages",
        data=json.dumps(body).encode(),
        headers={"x-api-key": KEY,
                 "anthropic-version": "2023-06-01",
                 "content-type": "application/json"},
    )
    with urllib.request.urlopen(req, timeout=120) as r:
        return json.load(r)["usage"]

for i in range(3):
    u = call(f"第 {i+1} 次:用一句话回答 1+1 等于几")
    read  = u.get("cache_read_input_tokens", 0)
    write = u.get("cache_creation_input_tokens", 0)
    plain = u.get("input_tokens", 0)
    total = read + write + plain
    rate  = read / total * 100 if total else 0
    print(f"第{i+1}次  未缓存={plain:6}  写入={write:6}  命中={read:6}  命中率={rate:5.1f}%")