AI model cost per task: synthetic examples / AI 模型每任务成本：合成算例

Run / 运行: python3 calculate.py
Python standard library only. No network requests, API calls, or paid services.
仅使用 Python 标准库，无联网、API 调用或付费服务。

All prices, token counts, outcomes, and toy benchmark weights are invented.
所有费率、token 数、成败与玩具评测权重均为人为设定，非真实模型评测或供应商现价。
Amounts: USD. Rates: USD per 1,000,000 native tokens.
金额：美元。费率：每百万原生 token 的美元价格。

Input buckets are disjoint: total_input = uncached_input + cache_write + cache_hit.
Output buckets are disjoint: output_total = reasoning + answer.
The invented tariff charges cache-write tokens at the cache-write rate INSTEAD OF the regular input rate.
输入互斥分桶；本合成价表的缓存写入费替代普通输入费，不额外叠加。
Real providers may use different pricing and usage fields. Check their actual bill.
真实供应商可能采用其他费率与 usage 字段，须以其账单规则核对。

No cache-write on retry; all 10,000 repeated cacheable tokens are assumed cache hits.
Failed first generation is assumed fully charged. All retry tokens are charged again.
重试假设 10,000 可缓存 token 全命中；失败首轮完整计费，重试再次计费。

Separate examples: a 10-task retry workload; a 10-task 8-success workload;
and a two-benchmark toy weighted mean. Do not combine their totals.
后面三例是彼此独立的账本，不合并其费用。
Tool calls, network, sandbox, cache storage duration, tax and human review excluded.
不含外部工具、网络、沙箱、缓存存储时长、税费及人工复核成本。

Files: calculate.py, attempts.csv, results.json, selection-template.csv, README.txt.
