25 条 AI 编程工具记录里有 8 条是开源项目：你在比的可能不是同一种东西
$9/月到 $100/月、6 条按用户计价、只有 1 条写出免费额度——把这 25 条记录的定价原文摆开，看编程助手这一类为什么最容易比错。
口径（2026-09-18，对 376 条已发布记录跑一次）：分类取 `primaryCategory === "coding"`（25 条）；「有数字」= `pricing.detail.zh` 出现 0-9；「标了免费档」= `pricing.hasFreeTier === true`；「提到免费」= 该串含「免费」或 free；「席位口径」= 含「用户/」或 per user；「写出额度」= 匹配 \d+\s*(次|分钟|小时|条|篇|字符|credits|tokens)；「等待名单类」= 含 等待名单 / 待定 / 定制 / contact。开源与 API 两个数取布尔字段 `openSource`、`hasApi`。价格区间是该分类说明里出现过的美元数字取并集（$9 来自 Bolt.new，$100 来自 Claude Code 的 Max 档）。复算：`npx tsx` 里 `import { tools } from "@/lib/tools"` 按这些规则数；「上一次提交距今多久」这类我们没入库，所以文中只给自查方法、不给数字。模型类数据基准日期：2026-09-15（数据集更新后此日期与正文数字一并复核）
先说结论：这一类混了三种产品，价格不可互比
primaryCategory = coding 的有 25 条。它们不是一类东西：
- IDE / 编辑器型订阅（Copilot、Cursor、Windsurf、Phind、Tabnine、Augment Code…）：按月费、按席位；
- 终端 / Agent 型（Claude Code、Devin、Aider 类、Replit Agent…）：常挂在另一个订阅里，或根本没价；
- 开源库与自建组件（BAML、Instructor、Outlines、Gradio、Streamlit…）：价格是 0，成本全在你要不要自己运维。
站内能直接对上这件事的数字是：25 条里 openSource 为真的有 8 条、hasApi 为真的 7 条。
也就是说，你把「Cursor $20/月」和「某个开源框架」放在一张表里比价格时，比的是两种完全不同的成本结构。
一、25 条的定价原文长什么样
| 统计 | 条数 |
|---|---|
| 记录总数 | 25 |
| 定价说明里有数字 | 15 |
| 标了「有免费档」 | 17 |
| 说明里提到「免费」 | 21 |
| 写了「/用户/月」这类席位口径 | 6 |
| 写了带数字的免费额度 | 1（v0「免费200次生成」） |
| 写了「等待名单 / 定制 / 待定」 | 5 |
| 出现过的美元价范围 | $9 – $100 |
值得盯的是那行 1：整整 25 条里只有 v0 写明了「免费 200 次生成」。 次数口径对写代码的人来说含义很具体——一次生成不等于一次修改，一次失败的补全也算一次； 所以这一档真正的读法是「200 次尝试」而不是「200 次改动」。 其余 24 条要么只写「免费」，要么写「免费版 / 按量付费」（Groq），要么写「免费 / 开源」（那一类连「档」都没有）。
二、三种成本结构，各自怎么算
1. 按席位：站内 6 条写了「/用户/月」，例如 GitHub Copilot 记的是「$10/月 / $19/用户/月」、 CodeRabbit 记的是「免费 / Pro $12/用户/月 / 企业版」。算团队成本时先问清评审席位与开发席位是否同价。
2. 挂在别处：Claude Code 的原文是「包含在Claude Pro $20/月 / Max $100/月」。
这类条目不能单独比：它的成本是「你为宿主订阅付的钱 ÷ 你在那里面写代码的比例」。
要拆细，用 /free-tools/llm-api-cost-calculator 把你真实的 token 用量乘一遍，别拿月费当单价。
3. 开源：8 条 openSource = true。0 元授权不等于 0 元成本，差别在：能不能离线、模型自己出钱、
维护者是谁、issue 多久回。这些字段我们的表里没有，所以这一类我们不给推荐，只给你一条自查路径：
先看有没有 1.0 以上的 release、再看上一次提交距今多久，两个都能公开查到。
三、换算成「每次被采纳的改动多少钱」
编程助手最容易被高估也最容易被低估，原因同一个：分母选错。月费 ÷（每天真正被采纳的补全数 × 30）， 而不是 ÷ 生成次数。探索性的、被删掉的、报错重试的，都不该进分母。多数人第一次这么算完， 会发现自己在为「用得没那么多的东西」付钱——这不是坏消息，这是终于知道该砍哪一档。
顺带把「按 token」类工具的成本对齐方法放在这里：/free-tools/token-counter 能估你贴进去的中文与英文各花多少 token，
同一个改动请求用中文描述与用英文描述，token 数不是同一回事。
四、这套表管不了什么（诚实边界）
- **我们不测补全质量，也不测它在你代码库里的可用度。**那需要你的仓库、你的语言栈，我们不替你猜。
- 5 条写「等待名单 / 待定」的记录（Devin 是「等待名单 / 企业定价待定」），这类条目会先于其他条目过期： 它们随时可能第一次公布价格。别把「我们记的是待定价」读成「它贵」。
- 定价串有保质期，日期在每条记录底部。
一句话版本
先分清楚你在比的是订阅、Agent 还是开源组件，再谈价格。 比错的代价不是多花 $10，是把「免费的库」当成「零成本」，而它其实是「用运维时间付款」。
常见问题
- 开源那 8 条为什么不给「推荐」？
- 因为决定它们值不值的三件事都不在目录能记的范围内：你的模型成本、你的运维时间、上游还在不在维护。我们能记的是「有 API 吗」「开源吗」这类布尔值，记不了「装上去要不要两个人月」。写推荐等于把你们的实施成本当成我们的判断，所以这一类我们只把原文摆出来，把你该问的三个问题列出来。
- 「免费 200 次生成」这种口径，为什么不换算成「大约多少次改动」？
- 因为换算系数完全取决于你的用法：有人一次生成直接用，有人平均试三次才留一次。给一个「大约 100 次改动」听着有用，实际是把我们的假设写成了你的预算。我们的做法是把口径讲清楚——次数、字符、token、分钟是四种不同的计量——让你用自己的数去乘，站内的计数器就是为这一步准备的。
- 为什么 Copilot 那条既有 $10/月 又有 $19/用户/月？
- 因为我们抄的是页面上并列的两档：个人档按月、商业档按席位，页面上它们就是并排写的。我们不合并成「$10 起」——那会让人以为商业档也是十块。这类「同一行里两种计量」的记录在全站有 67 条含「用户/」字样的条目，比价格高低更重要的是先看清它按什么计。要对照席位口径，把 Copilot 与 Cursor 的详情页并排开最快。
接着读
延伸阅读：这些条目可以直接开通
- v0免费200次生成 / Pro $20/月25 条里唯一写出免费额度的那条：「免费200次生成」。前往官网
Eight of the 25 AI coding records are open-source projects: you may not be comparing the same kind of thing
From $9/mo to $100/mo, six records priced per user, exactly one stating its free quota: all 25 coding price strings laid out, and why this category invites the worst comparisons.
Rules (2026-09-18, one pass over the 376 published records): the category is `primaryCategory === "coding"` (25 records); 'has a digit' means 0-9 appears in `pricing.detail.zh`; 'free tier' means `pricing.hasFreeTier === true`; 'mentions free' means the string contains 免费 or free; 'per seat' means 用户/ or per user; 'states a quota' matches \d+\s*(次|分钟|小时|条|篇|字符|credits|tokens); 'waitlist wording' means 等待名单 / 待定 / 定制 / contact. The open-source and API counts read the booleans `openSource` and `hasApi`. The price band is the union of dollar figures in that category's strings ($9 from Bolt.new, $100 from Claude Code's Max tier). Reproduce by importing `{ tools } from "@/lib/tools"` in `npx tsx`. We do not store commit recency, so the article gives that as a self-check method and publishes no number for it.
The short version first: this bucket holds three kinds of product, and their prices are not comparable
There are 25 records with primaryCategory = coding, and they are not one thing:
- IDE / editor subscriptions (Copilot, Cursor, Windsurf, Phind, Tabnine, Augment Code): monthly, per seat.
- Terminal / agent tools (Claude Code, Devin, Replit Agent and friends): usually attached to another subscription, or not priced at all.
- Open-source libraries and self-hosted components (BAML, Instructor, Outlines, Gradio, Streamlit): licence costs nothing, and the bill arrives as your own ops time.
The number in our data that makes this visible: 8 of the 25 have openSource = true and 7 have hasApi = true.
So putting 'Cursor $20/mo' in the same table as an open-source framework compares two different cost structures,
not two prices.
1. What those 25 pricing strings actually contain
| Measurement | Records |
|---|---|
| Total in category | 25 |
| Price string contains a digit | 15 |
| Marked as having a free tier | 17 |
| String mentions 免费 / free | 21 |
| Written as 'per user/month' | 6 |
| States a free quota with a number | 1 (v0: '免费200次生成') |
| Says waitlist / custom / TBD | 5 |
| Dollar figures seen | $9 – $100 |
Look at that 1. Across 25 coding tools, only v0 writes its free quota down, and it writes it as 200 generations. For a developer that wording matters: one generation is not one accepted change — a failed completion consumes one too. Read that tier as '200 attempts', not '200 edits'. The other 24 either say only 'Free', or '免费版 / 按量付费' (Groq), or '免费 / 开源' — the last group does not even have tiers.
2. Three cost structures, and how to compute each
Per seat. Six records state '/user/mo' — GitHub Copilot is recorded as '$10/月 / $19/用户/月', CodeRabbit as '免费 / Pro $12/用户/月 / 企业版'. Before budgeting a team, ask whether reviewer seats cost the same as developer seats; that single question usually moves the number more than the discount does.
Bundled elsewhere. Claude Code reads '包含在Claude Pro $20/月 / Max $100/月'. You cannot price-compare that
standalone: its cost is 'what you pay for the host subscription × the share of it you spend coding'. To split it
honestly, run your real token volume through /free-tools/llm-api-cost-calculator instead of treating a monthly fee
as a unit price.
Open source. Eight records carry openSource = true. A zero licence is not a zero cost: the differences are
whether it runs offline, who pays for the model, who maintains it, and how fast issues get answers. Our table does not
hold those fields, so for this group we offer no recommendation — only a check you can run: is there a release at 1.0
or above, and how many days since the last commit. Both are public.
3. Converting to 'cost per change you actually kept'
Coding assistants get overvalued and undervalued for the same reason: the wrong denominator. Use monthly fee ÷ (completions you actually kept per day × 30), not the number generated. Exploratory, deleted and retried completions do not belong in the denominator. Most people, computing this once, discover they are paying for something they use far less than they assumed — that is not bad news, it is finally knowing which tier to drop.
For the token-metered ones, /free-tools/token-counter shows what your prompt costs in Chinese versus English; the
same request described in the two languages is not the same number of tokens.
4. What this table cannot do (stated plainly)
- We do not benchmark completion quality or how well a tool behaves in your repository. That needs your codebase and your language stack, and we will not guess it.
- Five records say 'waitlist / TBD' (Devin is '等待名单 / 企业定价待定'). Those go stale faster than any other entry: a price may appear at any moment. Do not read 'we recorded no price' as 'it is expensive'.
- The strings expire; each record's verification date is printed at the bottom of its page.
One-sentence summary
First work out whether you are comparing a subscription, an agent or an open-source component — then talk about price. The cost of getting this wrong is not $10 a month; it is treating a free library as zero-cost when you are actually paying with your own ops time.
FAQ
- Why no recommendation for the eight open-source records?
- Because the three things that decide their value are outside what a directory can record: your model cost, your ops time, and whether upstream still maintains the project. We can store booleans like 'has an API' or 'open source'; we cannot store 'this will cost you two person-months to integrate'. Writing a recommendation would mean turning your implementation cost into our judgement, so for this group we publish the raw wording and the questions you should ask instead.
- Why not convert '200 free generations' into roughly how many accepted changes?
- Because the conversion factor depends entirely on your habit: some people keep the first generation, others retry three times before keeping one. Publishing 'roughly 100 accepted changes' sounds useful while writing our assumption into your budget. So we state the counting units instead — attempts, characters, tokens, minutes are four different meters — and let you multiply with your own numbers; the counters on this site exist for exactly that step.
- Why does the Copilot record show both $10/mo and $19/user/mo?
- Because we copied two tiers that the page lists side by side: an individual monthly price and a business per-seat price. We do not collapse them into 'from $10', which would imply the business tier is also ten dollars. Rows mixing two meters in one line are common — 67 records site-wide contain the 用户/ wording — and the counting basis matters more than which number is lower. Opening the Copilot and Cursor pages side by side is the fastest way to compare seat conventions.
Where to go from here
- v0Free 200 generations / Pro $20/moThe only record of the 25 that states a free quota: '免费200次生成' (200 free generations).Open the site