Grok 프롬프트 캐싱 과금 가이드

cached_tokens = floor(matched prefix length / 128) * 128

# WRONG: dynamic content at the start of system, so the prefix changes every time and never hits
messages = [
    {"role": "system", "content": f"Current time {datetime.now()}. You are an assistant." + LONG_INSTRUCTIONS},
    {"role": "user", "content": question},
]

# RIGHT: long instructions and tool definitions stay stable up front, dynamic content goes in the user message
messages = [
    {"role": "system", "content": LONG_INSTRUCTIONS},          # stable, will hit
    {"role": "user", "content": f"Current time {datetime.now()}. {question}"},  # volatile, goes last
]

import os
from openai import OpenAI

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

# The prefix has to be long enough: below the thousand-token range you get essentially nothing
LONG_SYSTEM = open("long_instructions.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("Summarize the key points", "call 1")   # cold start: cached is 0 or a tiny value
ask("Give me 3 keywords", "call 2")         # expect cached close to the prefix length

[call 1] input=8804 cached=128
[call 2] input=8804 cached=8704

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

r2 = client.responses.create(
    model="grok-4.6",
    previous_response_id=r1.id,          # send only the new turn; the upstream carries the history
    input=[{"role": "user", "content": "Follow-up"}],
    store=True,
)
print(r2.usage.input_tokens_details.cached_tokens)