大家好，我是孟健。
过去 3 个月，我一人协同多个 AI 写了 27 万行代码。
完成近 3000 次提交。
全年在各类 AI 工具上的订阅花费，大概 1.7 万元。
在这套一人公司的商业交付体系里，我选模型只看两笔账。
一个是 Token 计费成本。
另一个是真实工程排障的交付率。
10 月 23 日到 24 日，月之暗面正式发布了 Kimi K3.1 Pro。
正好我自研的内容分发工具「字流」里，压着一个陈年 bug。
之前换了几个模型都没修好。
这次我直接把任务丢给 K3.1 Pro 实测。
从敲下回车到复现、修改、验证完毕，全程约 9 分钟。
说实话我没想到，整个过程我一行代码都没写。
这篇我把它的基准跑分、API 成本账单，以及修 bug 和做前端落地页的完整工程链路，彻底拆开讲清楚。
先看跑分与算账
K3.1 Pro 采用混合专家架构，总参数规模达 2.8 万亿。
官方称它为 3 万亿级参数模型。
它每次激活约 1040 亿参数。
上下文窗口支持 100 万 Token，很长的代码库和文档能一次读进去。
官方承诺在 2026 年 11 月 4 日前，开源完整模型权重。
在第三方评测机构 Artificial Analysis 的 AA 榜上，K3.1 Pro 拿了 55 分。
全球排第四。
the-decoder 10 月 24 日那篇报道里提到，排在它前面的只有三个模型。
也就是 Claude Opus 5.5（58 分）、Claude Sonnet 5.5（56 分），以及还没开放 API 的 Muse Wonder Full（56 分）。
再看国产模型，小米 MiMo、通义千问、DeepSeek、GLM 都在 45 分上下。
K3.1 Pro 直接拉开了整整 10 分差距。
比上一代 K3 的 44 分，也涨了 11 分。
国产模型第一次这么近地贴着 Claude 跑。
AA 榜看的是综合分。
作为独立开发者，我更在意的是编程与命令行实操。
挑三个核心基准数据来看。
一个是测长时间连续写代码的 SWE Marathon v1.1，它跑了 51.7%。
直接位列榜首。
这个成绩超过了 Claude Opus 5.5 的 46.3%，也高过 GPT-6 Astra 的 42.5%。
测 AI 自己干活的 Agents' Last Exam，表现同样位列前排。
再看测真实终端受限任务的 Terminal-Bench 4.0，它拿到 65.7%。
上一代 K3 只有 13%。
这个成绩只比 Claude Opus 5.5 低了 0.7 个百分点。
但比 GPT-6 Astra 高出了 7.5 个百分点。
再算一下独立开发者最关心的 Token 账单。
OpenRouter 和官方定价里写着，K3.1 Pro 的 API 价格比前代降了大概三分之一。
国内标准版输入是 13 元每百万 Token，缓存命中 1.3 元。
输出是 68 元每百万 Token。
国内高速版输入 26 元，输出 136 元。
海外美元版这边，标准版输入 2 美元，缓存命中 0.2 美元，输出 10 美元。
高速版则是输入 4 美元，输出 20 美元。
你想想看，跑分领先真不等于全场景通吃。
理论基准与真实业务代码库之间，永远隔着一层工程边界。
回到刚才那个问题，它在真实项目里到底行不行？
下面直接看实测。
我自研的产品「字流」一年内累计提交了 192 次代码。
收件箱采纳过超 1300 条评论和私信。
写好一篇稿子，它能一键分发到各个平台。
其中 X 长文章这条链路，一直有个毛病。
一篇文章 11 张配图，自动插进去总会少一张。
永远只插进 10 张。
正文里还会留下一串乱码一样的占位文字。
每次我都得打开草稿手动补图、删字。
这事到底难在哪？
X 的编辑器有一套自己的光标位置记录机制。
程序在外部强行挪动光标，编辑器不一定认。
结果就是程序以为选中了占位文字，实际选中的却是别的东西。
更麻烦的是，这个问题只有在真实登录的 X 页面里才会触发。
平时写的自动化测试根本测不出来。
这种活光会写代码不够。
它得会读代码，会自己跑起来复现，还得会自己验证。
这次我用的是官方命令行编程工具 Kimi Code。
在项目目录里一行命令启动，直接调用 K3.1 Pro。
kimi -m kimi-code/k3.1-pro --yolo
这里的 --yolo 允许它读写文件和跑命令，不用每次停下来等我确认。
只有在自己的项目里，且改动都能用 git 追踪时，我才会这么开。
我给的提示词没有搞复杂的套路，只立了四条死规矩。
第一是现象，说明 11 张图只插进 10 张，还剩一个占位，附上日志原话。
第二是复现命令，给一条能直接跑起来的命令。
第三是修改范围，只许改指定的那一个文件，其它文件一律不准碰。
第四是红线，只存草稿，严禁点发布。
改完必须自己跑一遍验证，再用中文告诉我原因和结果。
9 分钟零人工介入排障
K3.1 Pro 接手后的第一步，最让我服气。
它先读了相关的 3 个文件。
它没有急着改代码，而是先把我的复现命令原样跑了一遍。
日志里当场抓到了现行。
程序本来要选占位文字，实际选中的却是一段正文标题。
很多 AI 拿到 bug 的第一反应，是先猜问题在哪，然后直接改。
K3.1 Pro 的顺序是对的。
先复现，再动手。
这就是有经验的工程师干活的样子。
定位到根因后，它的改法很干脆。
它不再从外部强行挪光标。
而是模拟鼠标点击占位文字，让 X 编辑器自己感知光标位置。
原来代码里「选中文字不对就不删」的校验保险，它也完整保留了下来。
改完代码，它立刻跑了一遍构建和复现命令。
它还顺手检查了 git 状态，确认只改了我允许的文件。
几分钟后，它给出三段总结，分别写明了根因、改法和验证结果。
正文图片 11/11 张，残留占位 0 个，全程没点发布。
我随后自己跑了一遍验证。
打开 X 草稿一段一段对着原稿核对。
11 张图全部归位，乱码占位文字一个不剩。
这个换了几个模型都没啃下来的硬骨头，9 分钟解决了。
修 bug 考的是工程逻辑，我更在意的是它的前端审美与交付闭环。
我让它给字流桌面版做一个落地页首屏，外加一屏三个卖点。
我给的约束很明确。
要有设计感，别做千篇一律的紫色渐变模板。
中文排版要讲究，双端都要做好响应式适配。
严禁编造用户数和评分这类假数据。
我还给它挂载了两个前端设计 skill，也就是 taste 和 impeccable。
它干活的流程很扎实。
先定视觉方向，写了一页产品说明，随后才开始写代码。
写完后，它自己调起浏览器，在桌面端和移动端两个尺寸下截图检查。
第一轮自检，它自己挑出 3 处毛病。
包括手机端面板掉出橙色区、说明没点明产品、第二屏界面尺寸太小。
它自己改完代码再查，直到检查全部通过。
它给出的设计思路，是以「中文活字校样台」为概念。
用石墨黑、纸白和信号橙替代模板化 AI 渐变。
直接把「一篇原稿分流到不同平台」做成主视觉。
首屏主标题定为「写一篇，各自成章。」
右侧面板可以直接点击切换小红书、知乎、抖音。
同一篇原稿会实时渲染成对应平台的排版。
第二屏的三个卖点配了界面示意，没堆砌廉价卡片。
切到手机尺寸时，布局会自动转为单列。
哪些坑你得提前知道
实测下来，K3.1 Pro 的工程能力确实能打。
但作为一人公司的主力工具，我觉得有几个客观短板必须说在前面。
第一个是超长难任务仍有差距。
在 AA 榜上，Claude Opus 5.5 比它高 3 分。
官方公布的长程编程评测里，它的得分从 K3 的 55.95% 升到了 61.2%。
但面对 GPT-6 Astra 依然落后，科学类的命令行任务差距更大。
官方也坦白，去掉网络攻防数据后，在安全测试上 pass@3 是 81.0%。
网络安全方向的能力，确实还落后于闭源前沿模型。
第二个是推理速度偏慢。
修那个 X 编辑器的 bug 用了 9 分钟。
做落地页加上多轮自检，前后跑了半个多小时，中间还断了一次连接。
如果你对响应速度敏感，建议切到同步发布的 Highspeed 版。
官方数据显示它输出能快 2 到 3 倍，但 API 成本也会翻倍。
第三个是套餐额度消耗极快。
这是我最想吐槽的一点。
修 bug 和做页面要跑好几轮测试与截图，Token 消耗量极大。
这次我刚准备测，Kimi Code 就提示本周额度用尽，要么等刷新要么加购。
10 月 26 日，官方账号 @月之暗面Kimi 也专门发文致歉。
承认会员使用时出现限流和不必要额度消耗，并重置了额度。
如果你是极度关注交付产出的独立开发者，我的结论很明确。
手头有拖了很久的复杂工程 bug，或者要做高品质前端界面，我非常推荐用它。
但要是做超大规模跨模块的长程重构，或者追求极高响应速度，建议先等等。
最后把这套排障提示词模板留给你，复制到项目里就能用。
1. 故障现象：[描述具体少了几张图/报了什么错，直接贴终端日志原话]
2. 复现命令：[给出一条可在当前目录直接执行的测试或构建命令]
3. 修改范围：[明确列出允许修改的文件路径，严禁触碰其它文件]
4. 验收红线：[明确禁止执行的操作如发布/删库，要求改完后自行跑一遍复现命令验证，并用中文输出根因、改法与验证结果]
下一阶段，我会在自己的出海产品矩阵里实测更多模型。
你最近让 AI 处理过最难缠的工程任务是什么？
欢迎在评论区交流。
👋 我是孟健，前腾讯 T11、前字节技术 Leader，现在全职做 AI 编程。
🔥 更多 AI 编程实战：
- GitHub：@mengjian-github
- 专栏：AI编程实战
- 这期的完整清单：评论区回复「编程实测」，我把领取方式发你
觉得有用？点赞+收藏 就是最大支持 🙏
本文由作者借助 AI 辅助整理，内容经作者实测与审校。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Hello, I'm Meng Ken.
Over the past three months, I've been working with multiple AIs writing 270 million lines of code.
Nearly 3000 submissions completed.
The annual subscription cost for various AI tools is approximately $17,000.
In this one-man business delivery system, I'm looking at only two books.
One is the Token billing cost.
Another is the delivery rate of real engineering barriers.
From 23 to 24 October, the dark side of the moon officially became Kimi K3.1 Pro.
It's just that I've been studying my own text distribution tool, "Trips", and I've been working on it for a long time.
There were several models that were not fixed.
This time I dropped the task directly to K3.1 Pro.
From knock-back to recurrence, modification, validation, approximately 9 minutes.
Honestly, I didn't think I'd write a line of code for the whole thing.
This piece I'm going to break down its benchmark run, the API cost bill, and the complete engineering link to fix the bug and do the front page.
Let's see the score and the score.
K3.1 Pro uses a hybrid expert structure with a total parameter size of 2.8 trillion.
It's officially called a 3 trillion-billion-degree parameter model.
It activates about 104 billion parameters each time.
context window supports 1 million Token, with a long code library and documentation that can be read at once.
Officially committed to complete model weights by November 4, 2026.
K3.1 Pro took 55 points on the AA list of the third-party rating agency Artificial Analysis.
The world ranks fourth.
In that story on October 24th, the decoder said that there were only three models ahead of it.
This is Claude Opus 5.5 (58 minutes), Claude Sonet 5.5 (56 minutes) and Muse Wonder Full (56 minutes) which is not open to API.
And look at the national model, Mimo, Qwen, DeepSeek, GLM, all up and down 45 points.
K3.1 Pro directly closes the 10-point gap.
44 points higher than the previous generation K3, and 11 points higher.
For the first time, the national model was so close to Claude.
The AA list looks at composite points.
As an independent developer, I care more about programming with command line.
Three core baseline data are selected.
One is the SWE Marathon v1.1, a continuous code measuring a long period of time, which runs 51.7 per cent.
Direct top of the list.
This was more than 46.3 per cent of Claude Opus 5.5 and 42.5 per cent of GPT-6 Astra.
Acts' Last Exam for AI to work on its own, showing the same rank in the front row.
It received 65.7 per cent of Terminal-Bench 4.0 for the real terminal restricted mission.
The K3 of the previous generation was only 13%.
This is only 0.7 percentage points lower than Claude Opus 5.5.
But it's 7.5 percentage points higher than GPT-6 Astra.
And calculate the Token bill, which is the main concern of independent developers.
OpenRouter and official pricing says that K3.1 Pro's API price is about a third lower than the previous generation.
Domestic standard input is $13 per million Token, cache hit 1.3.
The output is $68 per million Token.
Domestic high-speed input of $26, output of $136.
On this side of the overseas dollar version, the standard version entered $2.00, cache hit 0.2.
The high-speed version is for $4 and $20.
Think about it, running ahead is not the same as eating the whole scene.
There is always an engineering boundary between the theoretical benchmark and the real business code library.
Go back to the question. Is it working on the real project?
We'll see what happens next.
My self-researched product "Typography" submitted 192 codes in one year.
Over 1,300 comments and private letters were included in the inbox.
Write a piece that can be distributed to the platforms.
There's a long article on this link, and there's always a problem.
One article, 11 maps, will always be missing.
You'll never fit in ten.
In the text, there will be a bunch of uncoded places.
Every time I have to open a draft of a manual supplement and delete.
What's so hard about this?
X's editor has its own cursor location recording mechanism.
The programme is forced to move the cursor outside, which is not necessarily recognized by the editor.
The result is that the program thought that the placeholder text had been selected and that what was actually selected was something else.
Even more troublesome, this problem will only be triggered on the X page that is actually logged in.
Automation tests normally written cannot be detected.
It's not enough to write code.
It's got to read the code, run it back and verify it.
This time I used the official command line programming tool Kimi Code.
Starts with a line of commands in the project directory, directly calling K3.1 Pro.
kimi -m kimi-code/k3.1-pro --yolo
Here-yolo allows it to read and write documents and run orders without having to stop for confirmation.
I only do this when I'm in my own project, and when all changes can be tracked with guit.
The prompt I gave you didn't have a complicated set of rules.
The first is the phenomenon, which means that 11 pictures were inserted in only 10, leaving one place with the original log message.
The second is to repeat the order and give an order to run straight up.
The third is to modify the scope of only one document that has been assigned and none of the other documents will be touched.
The fourth is the red line, which has only drafts and is strictly restricted to publish.
This must be done by running it through and then telling me why and the results in Chinese.
Nine minutes of zero manual intervention.
K3.1 Pro takes over the first step, and I'm the best at it.
It reads three relevant files first.
Instead of rushing to change the code, it ran out of my duplicate order.
The log caught the current one on the spot.
The program would have chosen the placeholder text, but actually the text title.
A lot of AI gets bug's first reaction is to guess where the problem is and then change it directly.
K3.1 Pro is in the right order.
Come on, come on.
That's what experienced engineers do.
Once it's positioned to the root cause, it's very simple.
It no longer forced the cursor from outside.
It's a simulator of mouse clicked placed text that allows the X editor to sense the cursor position.
The check-up policy for "selecting the wrong text without deleting it" in the original code was also kept intact.
After the code was changed, it immediately ran over the construction and reproduction orders.
It also checked the guit state, confirming that only the documents I allowed had been altered.
A few minutes later, it gave three paragraphs summarizing the causes, the modification and the results of the validation.
11/11 images, left zero, no publish.
I ran over it myself.
Opens a paragraph of the draft X to check the original.
11 All the graphs are in place, and there's no placeless text left.
It's a hard bone that hasn't eaten out of several models. It's been solved in nine minutes.
Fixing bugs is engineering logic, and I care more about its front-end aesthetics and delivery rings.
I have made it a front page screen for the script desktop, plus a screen of three sales points.
My constraints are clear.
You need to have a sense of design. Don't do a whole purple gradient template.
Chinese scripts are very well written, and both sides are responsive.
It is strictly forbidden to compile and rate such false data.
I also mounted it with two front-end designs skill, that's why it is.
It works well.
The visual orientation was followed by a one-page product description, followed by code writing.
After writing, it pulls up its own browser and checks the screenshots below the desktop and moving ends.
First round of self-censorship, it singled out three problems.
These include the release of the end panel of the mobile phone from the orange area, the lack of an indication of the product, and the small size of the second screen interface.
It changed its own code and checked it until all checks passed.
It gives a design idea based on the concept of a Chinese-language proofshop.
Replace the templates with graphite black, paper white and signal orange.
It is a direct visualization of "a copy of the original to different platforms".
The main title of the screen is "Current one, separate chapters."
Right panel can click to switch REDnote, Zhihu, Douyin.
The same manuscript will be printed in real time into the layout of the corresponding platform.
The three sales points of the second screen were marked with interfaces and no cheap cards were stacked.
When you cut a cell phone size, the layout automatically changes to a separate column.
You need to know which pits.
K3.1 Pro's engineering capability does work.
But as the backbone of a single company, I think there are a few objective shortboards that must be put ahead.
The first is that there is still a gap in the extremely difficult task.
On the AA list, Claude Opus 5.5 points higher than it by three.
In the official long-range programming evaluation, its score rose from 55.95 per cent of K3 to 61.2 per cent.
But the GPT-6 Astra is still lagging behind, and the gap between the scientific command line tasks is even greater.
Officially, it was 81.0 per cent on security tests for pass@3.
The ability to steer cybersecurity does lag behind the closed-source frontier model.
The second is the slow pace of reasoning.
The bug for repairing the X editor took nine minutes.
The landing page was accompanied by multiple self-inspection rounds, which ran for more than half an hour and cut off one connection.
If you are sensitive to the speed of response, it is recommended to switch to the publish-Sync Highspeed version.
Official data show that it can output 2-3 times faster, but API costs double.
The third is the quota, which consumes very quickly.
That's the last thing I want to say.
Fixing bugs and making pages involves several rounds of testing and screenshot, and Token consumes a lot.
I was just about to test, and Kimi Code was suggesting that this week quota ran out, either waiting to be refreshed or bought.
On October 26, the official account number @The Dark Side of Moon Kimi also sent a special apology.
rate limit and unnecessary quota consumption were admitted and quota was replaced.
If you are an independent developer with great interest in delivering outputs, my conclusions are clear.
There's a long delay in complex engineering bugs, or high quality front-end interfaces, which I highly recommend.
However, if a long-range re-engineering of super-large-scale cross-modules is to be done, or if a very high response rate is to be pursued, it is recommended, etc.
I'll leave you with the prompt template, which will be copied into the project.
1. Fragmentation: [Description of specific missing graphs/relationships and direct posting of terminal log words]
2. Recurring command: [Give a test or build command that can be executed directly in the current directory]
3. Scope of change: [Explicitly list the document paths that allow for modification and ban contact with other documents]
4. acceptance check red line: [Actions expressly prohibited from execution, such as publish/Deleting Library, require self-reprinting and re-validation commands once they have been completed, and output root causes, modification and validation results in Chinese]
In the next phase, I'm going to measure more models in my own offshore matrix.
What is the most difficult engineering mission you've ever asked AI to handle?
An exchange in the comment area is welcome.
I'm Meng Ken, ex-Tencent T11, ex-byte technology Leader, now full time doing AI programming.
🔥 More AI programming field:
- GitHub：@mengjian-github
- Column: AI programming field
- A complete list of this issue: the comment section returns "programming" and I'll send you the collection.
You think it works? It's the greatest support for you.
The text is organized by the author with the help of AI, and is measured and revised by the author.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.