Prompt Cache 片段：命中率不够就别指望折扣
只打印缓存相关字段。折扣只发生在 cached_tokens（cache read）。前缀太短、有改动、第一次写入、或命中率太低，都不要按缓存单价估算。
本页不发起请求；不收 Key；中转 curl 不在本站。
← 计费示例 · 计费路径 · 官方 usage 文档
输入input_tokens_details.cache_write_tokens
输出—
缓存input_tokens_details.cached_tokens
hit_rate · 门槛：新模型约 1024 tokens 前缀，旧模型约 2048。低于门槛、前缀有任何改动、或首次请求，折扣都不成立。
import os
from openai import OpenAI
# Prefix must be byte-identical across calls. Short prefixes never hit.
# Newer models ~1024 tokens minimum; older ~2048. First call is a write, not a discount.
STABLE_PREFIX = ("You are a billing ledger. Keep answers short. ") * 160
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
response = client.responses.create(
model="gpt-4o-mini",
input=[
{"role": "system", "content": STABLE_PREFIX},
{"role": "user", "content": "Reply with one word: ok"},
],
)
usage = response.usage
input_tokens = usage.input_tokens if usage else 0 # 输入
output_tokens = usage.output_tokens if usage else 0 # 输出
details = usage.input_tokens_details if usage else None
cached = (details.cached_tokens or 0) if details else 0 # 缓存（read）
write = (getattr(details, "cache_write_tokens", 0) or 0) if details else 0
hit_rate = (cached / input_tokens) if input_tokens else 0
print("输入", input_tokens)
print("输出", output_tokens)
print("缓存", cached)
print("cache_write", write)
print("hit_rate", round(hit_rate, 4))
# Discount does NOT apply when:
# - prefix shorter than the cacheable minimum
# - prefix bytes change (whitespace counts)
# - this is the first request (write, not read)
# - hit_rate is too low to beat uncached input $/M
| 字段 | 对上列 | 说明 |
|---|---|---|
input_tokens_details.cached_tokens | 缓存 | 唯一按缓存读 $/M 的字段。命中率 = cached / 输入。 |
input_tokens_details.cache_write_tokens | 输入 | 写入不是折扣。接近输入价，有的模型还略贵。不要记进缓存列。 |
hit_rate | — | 门槛：新模型约 1024 tokens 前缀，旧模型约 2048。低于门槛、前缀有任何改动、或首次请求，折扣都不成立。 |