埃森哲Accenture
公司中央每人每月只出 29 美元，超支挂项目预算。此前重度用户每月烧 500 到 1000 美元，多是编程 Agent 用户。Central funding covers $29 per person per month; overages go to project budgets. Heavy users previously burned $500 to $1,000 a month, mostly on coding agents.
AI 成本治理工具包AI Cost Governance Toolkit
三部分一套工具，点页签切换：一、预算规划，按使用档位算清账，并把「公司承担」和「组织总成本」分开看；二、模型选型工作表，月成本随填随算；三、限流与降级清单，逐项对照落地。所有输入只在你的浏览器里处理。One toolkit in three parts, switched by the tabs below: (1) a budget planner by usage tier that separates central company cost from true total cost; (2) a model selection worksheet whose monthly cost updates as you type; (3) a rate-limit and degradation checklist to work through item by item. Everything you enter is processed only in your browser.
组成部分Three parts预算规划 · 模型选型 · 限流与降级Budget planning · Model selection · Rate limits and fallbacks
成本口径Cost basis公司承担与组织总成本分开算Company borne cost and total organisational cost, counted separately
选型工作表Selection worksheet月成本随填随算Monthly cost, calculated as you type
免费使用这件工具Use this tool free
留邮箱即可免费使用全部工具。注册一次，之后本站的工具直接打开，不用再填。Leave your email to use every tool free. Register once, and the tools on this site open directly afterwards, with nothing more to fill in.
验证码已发到 ，请查收邮件，把 6 位验证码填到下面。10 分钟内有效，没收到先看垃圾箱。A code was sent to . Enter the 6 digit code from the email below. It expires in 10 minutes. If it is not in your inbox, check the spam folder.
提交没有成功，请检查网络后重试。The submission did not go through. Please check your connection and try again.
为确认邮箱是真人真箱，注册前会先往邮箱发一封验证码邮件，输码后即完成注册。To confirm the address belongs to a real inbox, registration starts with a verification code sent to your email. Entering the code completes registration.
邮箱只用于发送资料与通知，可随时退订。邮箱怎么处理，见隐私政策。Your email is used only to send materials and notices, and you can unsubscribe at any time. See the Privacy Policy for how email is handled.
一、预算规划Part 1 · Budget Planner
按档位填人数和人均月支出，设每人封顶，再选超出部分由谁付，结果实时更新。Enter headcount and per-person monthly spend by tier, set a per-person cap, then choose who pays for anything over the cap. Results update live.
按使用档位填写（人数 × 人均每月支出）By usage tier (headcount × monthly spend per person)
轻度Light估算假设，非行业平均Planning assumption, not an industry average
中度Moderate估算假设，非行业平均Planning assumption, not an industry average
重度Heavy估算假设，非行业平均Planning assumption, not an industry average
失控Unmanaged估算假设，非行业平均Planning assumption, not an industry average
每人每月封顶（美元）Monthly cap per person ($)
超出部分由谁付Who pays for anything over the cap
各档人均支出（50 / 200 / 800 / 2000 美元）是估算假设，用来做规划测算，不是行业平均数，可按你团队的真实账单改。重度用户的量级锚定公开报道：埃森哲、Uber 披露的重度用户每月约 500 至 2000 美元。The per-tier spends ($50 / $200 / $800 / $2,000) are planning assumptions for estimation, not industry averages: edit them to match your real bills. The heavy-tier magnitude is anchored to public reporting: Accenture and Uber have disclosed heavy users at roughly $500 to $2,000 per month.
二、模型选型工作表Part 2 · Model Selection Worksheet
把候选模型按同一套价格口径摆在一起：填入价、输出价、缓存价、月调用量与平均输入输出，月成本随填随算。价格务必按厂商官网核对后再写入决策。Line candidate models up on one pricing basis: enter input, output and cache prices, monthly calls and average input and output, and the monthly cost updates as you type. Always verify prices against the vendor's official pricing before deciding.
| 模型Model | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($) | 备注Notes | 删行Delete |
|---|---|---|---|---|---|---|---|---|---|
| 模型Model示例行Example row | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($)$0.00 | 备注Notes | |
| 模型Model | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($)$0.00 | 备注Notes | |
| 模型Model | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($)$0.00 | 备注Notes | |
| 模型Model | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($)$0.00 | 备注Notes | |
| 模型Model | 输入价（美元/百万 token）Input price ($/1M tokens) | 输出价（美元/百万 token）Output price ($/1M tokens) | 缓存价（美元/百万 token）Cache price ($/1M tokens) | 月调用量（次）Monthly calls | 平均输入（token/次）Avg input (tokens/call) | 平均输出（token/次）Avg output (tokens/call) | 月成本（美元）Monthly cost ($)$0.00 | 备注Notes | |
| 月成本合计Total monthly cost | $0.00 | ||||||||
月成本口径：月成本 = 月调用量 ×（平均输入 × 输入价 + 平均输出 × 输出价）÷ 1,000,000。缓存价只作比价参考，未计入月成本：各家缓存命中口径不同，要计入先定命中率，再另行折算。第一行是示例行，示例价格仅示意。Monthly cost basis: monthly cost = monthly calls × (avg input × input price + avg output × output price) / 1,000,000. The cache price is a comparison reference only and is not included: cache-hit definitions differ by vendor, so set a hit rate first and convert separately. The first row is an example row; its prices are illustrative only.
三、限流与降级清单Part 3 · Rate-Limit & Degradation Checklist
逐项对照你团队的现状：已经落地的勾上，没做到的就是下一步。勾选只在本页计数，刷新不保留。Work through it against your team's current setup: tick what is already in place; the rest is your to-do list. Ticks are counted on this page only and are not kept after a refresh.
依据：埃森哲自 2026 年 10 月 1 日起中央每人每月只出 29 美元，超支挂项目 WBS，重度用户每月 500 至 1000 美元，曾有账号因用量极端被停用；Uber 2026 年全年 AI 编程预算四个月烧完，2026 年 6 月起每人每月每款工具封顶 1500 美元，并上线内部用量仪表盘与排行榜。Basis: Accenture: from 1 Oct 2026, central funding of $29 per person per month, overages charged to project WBS, heavy users at $500 to $1,000 a month, and one account deactivated for extreme use. Uber: its full-year 2026 AI coding budget burned in four months; from June 2026 a $1,500 cap per person per month per tool, plus an internal usage dashboard and leaderboard.
已落地 In place: 0 / 15
限额Caps已落地 0 / 40 / 4 in place
告警Alerts已落地 0 / 40 / 4 in place
熔断与降级Circuit-breaking & degradation已落地 0 / 40 / 4 in place
归属与复盘Attribution & review已落地 0 / 30 / 3 in place
真实公司的数字Real companies, real numbers
UberUber
2026 年全年的 AI 编程预算四个月烧完，个别工程师每月 500 到 2000 美元。之后每人每月每款工具封顶 1500 美元。Its full-year 2026 AI coding budget was gone in four months, with individual engineers at $500 to $2,000 a month. The fix: a $1,500 cap per person per month per tool.
为什么 AI 账单天生不封顶Why the AI bill never caps itself
- 1Agent 不是聊天。聊天按句付钱，Agent 自己循环干活、反复读上下文、还能拉起一群子 Agent，token 是乘法不是加法。Agents are not chat. Chat charges per message; agents loop on their own, re-read context, and spawn sub-agents. Tokens multiply, they do not add.
- 2预算是按旧世界估的。很多公司的 AI 预算定于编程 Agent 普及之前，参照的是聊天时代的用量。Budgets were sized for the old world. Many AI budgets were set before coding agents, using chat-era usage as the baseline.
- 3没人盯，就没人省。用量没有归属、没有仪表盘，账单只会涨不会停。Nobody watching means nobody saving. With no cost attribution and no dashboard, the bill only goes up.
大公司在用的五招Five moves the big companies use
- 1封顶：按人、按工具设每月上限（Uber 式）。Cap it: a monthly limit per person, per tool (the Uber move).
- 2归属：超支必须挂到具体项目，花钱要找得着主（埃森哲式）。Attribute it: overages must land on a specific project budget (the Accenture move).
- 3分级：普通档和高级档分开，额度分层。Tier it: separate standard and premium tiers with different allowances.
- 4复审回收：许可证和额度定期复审，不用的收回。Review and reclaim: licenses and allowances are reviewed regularly; unused ones are taken back.
- 5监控加例外：仪表盘盯用量，超额可以申请，但要过审。Monitor with exceptions: a dashboard watches usage; overages need an approved exception.