Token 计费与上下文长度：被忽视的成本杀手

#主流模型定价对比（单位：USD per 1M tokens）
models = {
    "GPT-4o (2024-05)": {"input": 2.5, "output": 10.0},
    "GPT-4o-mini":           {"input": 0.15, "output": 0.60},
    "Claude Sonnet 4":       {"input": 3.0, "output": 15.0},
    "Claude Haiku 3.5": {"input": 0.8, "output": 4.0},
    "Qwen2.5-72B (API)":     {"input": 0.5, "output": 1.5},
    "DeepSeek V3":           {"input": 0.27, "output": 1.1},
}

def calc_cost(model, input_tokens, output_tokens):
    return (input_tokens * models[model]["input"] + output_tokens * models[model]["output"]) / 1e6

print("单次调用成本估算（100 input + 500 output tokens）：")
for model, prices in models.items():
    cost = calc_cost(model, 100, 500)
    print(f"  {model:<25}: ${cost:.4f}")

print("\n日均 10000 次调用的月成本：")
for model, prices in models.items():
    daily = calc_cost(model, 100_000, 5_000_000)
    monthly = daily * 30
    print(f"  {model:<25}: ${monthly:.2f} / 月")

import numpy as np

def effective_context_ratio(actual_tokens, max_tokens):
    usage_ratio = actual_tokens / max_tokens
    if usage_ratio < 0.3:
        return 0.95
    elif usage_ratio < 0.5:
        return 0.85
    else:
        middle_loss = (usage_ratio - 0.5) * 0.4
        return max(0.7, 0.85 - middle_loss)

print("有效上下文比例估算：")
for max_ctx in [32, 128, 200]:
    print(f"\n  {max_ctx}K 上下文窗口:")
    for actual in [8, 32, 64, 128]:
        if actual > max_ctx:
            continue
        ratio = effective_context_ratio(actual, max_ctx)
        print(f"    {actual}K实际使用:有效率约 {ratio*100:.0f}% (利用率 {actual/max_ctx*100:.0f}%)")

print("\n结论: 128K 窗口用到 64K 时，中间信息质量已显著下降")

# 使用 tiktoken 计算 token 数（GPT 同款分词器）
import tiktoken

enc = tiktoken.get_encoding("cl100k_base")

texts = [
    "Running shoes for men breathable lightweight",
    "男士轻便透气跑步运动鞋",
    "Nike Air Max 270, the best-selling sneaker with Air Max cushioning",
    "Nike Air Max 270 超轻缓震跑步鞋男款黑武士透气网面",
    "USD 299.99",
    "$299.99",
]

print("Token 数对比（Tiktoken cl100k_base）：")
for text in texts:
    tokens = enc.encode(text)
    print(f"  {len(tokens):>3} tokens | {text[:50]}")

# 关键观察：
# - 英文品牌名 "Nike" = 1 token
# - 中文字符串 token 数普遍较高
# - 数字+货币符号两种写法 token 数相近

def route_task(task, model_pool):
    if task["complexity"] == "low":
        return min(model_pool, key=lambda m: m["cost_factor"])
    elif task["complexity"] == "medium":
        return sorted(model_pool, key=lambda m: m["cost_factor"])[1]
    else:
        return max(model_pool, key=lambda m: m["capability"])

model_pool = [
    {"name": "GPT-4o-mini", "capability": 0.6, "cost_factor": 0.1},
    {"name": "Claude Haiku", "capability": 0.65, "cost_factor": 0.12},
    {"name": "GPT-4o", "capability": 0.95, "cost_factor": 1.0},
    {"name": "Claude Sonnet 4", "capability": 1.0, "cost_factor": 1.2},
]

tasks = [
    {"task": "产品标题分类", "complexity": "low"},
    {"task": "竞品分析报告生成", "complexity": "high"},
    {"task": "多语言翻译", "complexity": "medium"},
]

for task in tasks:
    best = route_task(task, model_pool)
    print(f"  {task['task']} ({task['complexity']}) -> {best['name']}")

# 精简 System Prompt 的前后对比
original_system = """
你是一个专业的跨境电商产品经理助手。

你的专长：
- 分析亚马逊、Temu、SHEIN 等平台的产品标题和描述
- 提炼产品的核心卖点、目标人群、适用场景
- 识别标题中的关键属性（品牌、型号、颜色、尺码、材质）
- 提供多语言产品本地化建议（英语、西班牙语、德语、法语）

输出格式要求：
- 输出必须严格遵循用户指定的格式，不额外发挥
- 如果信息不足，明确说"信息不足，无法判断"
- 涉及价格、销量等数据时，标注"未提供"而不是编造

语言风格：专业、简洁、直接，关键结论放在最前面
"""

精简_system = """你是跨境电商产品分析助手。
专长：产品卖点提炼、属性识别、多语言本地化。
规则：严格按指定格式输出；信息不足说"未提供"；不编造数据。
风格：专业简洁，结论先行。"""

def estimate_tokens(text):
    return len(text.split()) * 1.3

orig_tokens = estimate_tokens(original_system)
mini_tokens = estimate_tokens(精简_system)
saving_pct = (orig_tokens - mini_tokens) / orig_tokens * 100

print(f"原始 System Prompt: ~{orig_tokens:.0f} tokens")
print(f"精简后: ~{mini_tokens:.0f} tokens")
print(f"节省: {saving_pct:.0f}% tokens")
print(f"月均万次调用节省: ${(orig_tokens - mini_tokens) * 10000 * 2.5 / 1e6:.0f}")

def rag_cost(num_chunks, chunk_avg_tokens, model="GPT-4o"):
    input_tokens = 50
    retrieved_tokens = num_chunks * chunk_avg_tokens
    system_tokens = 200
    total_input = input_tokens + retrieved_tokens + system_tokens
    output_tokens = 300
    return (total_input * 2.5 + output_tokens * 10) / 1e6

print("RAG 成本对比（GPT-4o）：")
configs = [
    (1, 500, "只召回 top-1，chunk 长"),
    (3, 300, "top-3 中等 chunk"),
    (5, 200, "top-5 短 chunk"),
    (10, 100, "top-10 极短 chunk"),
]

for num, avg, desc in configs:
    cost = rag_cost(num, avg)
    print(f"  {desc}: ${cost:.4f}/次 = ${cost*10000:.2f}/万次")

def batch_cost(num_items, system_tokens=200, per_item_tokens=50, output_per_item=100):
    single_total = num_items * (system_tokens + per_item_tokens + output_per_item)
    batch_total = system_tokens + num_items * (per_item_tokens + output_per_item)
    return single_total, batch_total

for n in [5, 10, 20, 50]:
    single, batch = batch_cost(n)
    saving = (single - batch) / single * 100
    print(f"  {n} 条任务: 逐条 {single} tokens -> 批量 {batch} tokens, 节省 {saving:.0f}%")