大模型API成本优化指南 — 5招节省80%的AI开支

import hashlib
import json

cache = {}

def cached_chat(client, model, messages, **kwargs):
    # 生成缓存key
    cache_key = hashlib.md5(
        json.dumps({"model": model, "messages": messages}, sort_keys=True).encode()
    ).hexdigest()

    # 命中缓存则直接返回
    if cache_key in cache:
        return cache[cache_key]

    # 未命中则调用API
    response = client.chat.completions.create(
        model=model, messages=messages, **kwargs
    )

    # 存入缓存
    cache[cache_key] = response
    return response

# ❌ 低效：逐条调用
for text in texts:
    result = client.chat.completions.create(
        model="deepseek-v4-flash",
        messages=[{"role": "user", "content": f"分类：{text}"}]
    )

# ✅ 高效：批量调用
batch_result = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{
        "role": "user",
        "content": f"请对以下文本进行分类，返回JSON数组：\n{json.dumps(texts, ensure_ascii=False)}"
    }],
    max_tokens=2048
)

def smart_chat(client, messages, complexity="auto"):
    """根据任务复杂度智能选择模型"""
    if complexity == "auto":
        # 简单启发式：根据消息长度判断
        total_len = sum(len(m["content"]) for m in messages)
        complexity = "simple" if total_len < 200 else "complex"

    model_map = {
        "simple": "deepseek-v4-flash",   # ¥2.02/百万Token
        "medium": "deepseek-v4",          # ¥25.06/百万Token
        "complex": "qwen3-5",            # ¥43.20/百万Token
    }

    model = model_map.get(complexity, "deepseek-v4")
    return client.chat.completions.create(
        model=model, messages=messages
    )