Grok Prompt Caching 快取計費指南

命中量 = ⌊匹配上的字首長度 ÷ 128⌋ × 128

# ❌ 錯：動態內容拼在 system 開頭，字首每次都變，永遠命中不了
messages = [
    {"role": "system", "content": f"當前時間 {datetime.now()}。你是一個助手。" + 長指令},
    {"role": "user", "content": 問題},
]

# ✅ 對：長指令和工具定義在前保持穩定，動態內容放 user 訊息裡
messages = [
    {"role": "system", "content": 長指令},          # 穩定，會命中
    {"role": "user", "content": f"當前時間 {datetime.now()}。{問題}"},  # 易變，放最後
]

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APIYI_API_KEY"],
    base_url="https://api.apiyi.com/v1"
)

# 字首要足夠長：千 token 量級以下基本吃不到快取
LONG_SYSTEM = open("long_instructions_zh.txt").read()


def ask(question: str, label: str):
    r = client.chat.completions.create(
        model="grok-4.6",
        messages=[
            {"role": "system", "content": LONG_SYSTEM},
            {"role": "user", "content": question},
        ],
    )
    cached = r.usage.prompt_tokens_details.cached_tokens
    print(f"[{label}] input={r.usage.prompt_tokens} cached={cached}")


ask("請概括要點", "第1次")    # 冷啟動，cached 為 0 或極小值
ask("請給3個關鍵詞", "第2次")  # 期望 cached 接近字首長度

[第1次] input=8804 cached=128
[第2次] input=8804 cached=8704

r1 = client.responses.create(
    model="grok-4.6",
    input=[{"role": "system", "content": LONG_SYSTEM},
           {"role": "user", "content": "第一個問題"}],
    store=True,
)

r2 = client.responses.create(
    model="grok-4.6",
    previous_response_id=r1.id,          # 只傳新增的一句，歷史由上游接續
    input=[{"role": "user", "content": "追問"}],
    store=True,
)
print(r2.usage.input_tokens_details.cached_tokens)