今天终于被 Codex 气到了！
这个“小辣鸡”一直吹的“无敌电脑操作能力”彻底崩溃了！
然后问题还没解决，配额又被清空了。 接下来还要合并 Codex 和 ChatGPT 配额，这还怎么玩啊！
我就今天用了一下下，只剩 48% 了！ 你说这 Plus 账号还有什么用呢？
我用的还不是 Astra，是 Sol。
我明显感觉现在这个配额比之前又少了很多！
不说配额了，说说“电脑操控”这个事情吧。
这个功能本来是很有用的，可以节省我们大量的时间和精力。你只要给一个任务，它自动帮你“用电脑”，非常爽。
最近刚好有一个需求，我就对比了几个 Agent 工具，发现同样是操作电脑，大家的差异还是很大的。
这次的测试对象主要有 ZCode（选 5.3 Flash）、Codex（选 GPT-6.1 Sol）、KimiCode（选 K3.1 Pro）。
全部都是用官方的智能体配上官方的模型！
我的测试场景是这样的：
Kimi 官网提供了一个“3D 建模”功能效果还不错。我希望它们能帮我读取本地的一批参考图，然后自动调用网页的功能，自动上传参考图，自动填写提示词，自动下载生成的文件，自动整理归档。
这个工作人工做非常繁琐，还要一直等着，让 Agent 做就非常合适。
花了大半天功夫，已经测完了，建模文件也拿到了。给大家分享一下三个工具的实战能力，同时包括时间消耗，费用等信息。
1、KimiCode
KimiCode 整体来说没啥大毛病，还可以更快一点，更智能一点！
它的大致执行逻辑如下：
我先查看 concepts\v2 里的图片，确认哪三张是人物及其文件名，再逐张上传到 Kimi，使用 Tripo3D 多视图生成、跟踪进度并下载回对应目录。
按最新生成时间，三个人物是 孙策、夏侯惇、甘宁。我会使用 views 里的三视图版本，模型保存到 pubg\docs\Tripo3D.glb\v2，与现有 V1 模型目录对应。
参考图已确认并归档。接下来我会在 Chrome 中新建 «三国 V2 Tripo3D 生成» 标签组，逐个提交任务；这些标签页会保留用于跟踪，你也可以随时让我关闭。
扩展的直接上传权限未开启，我改用电脑控制打开 Windows 文件选择器上传，不需要你修改浏览器设置。
孙策三视图已上传，提示词已填写。现在提交 Tripo3D 生成任务，并观察是否正常进入裁图和建模流程。
孙策任务已正常进入三视图裁分与 Tripo3D 生成流程。为减少等待，我会在独立标签页中继续提交夏侯惇和甘宁，再统一跟踪三个任务。
三个人物任务均已提交成功，正在后台生成。我会轮流检查进度，完成后立即下载到 V2，不需要你继续操作。
孙策已生成完成，GLB 内含 4K 贴图，约 51 MB。我先下载孙策，另外两个人物继续生成。
可以看到整个执行过程思路还是很清晰的！
基本上没有遇到啥问题，唯一的问题是上传权限，然后立马改用 Windows 文件选择器上传了！
它在 Kimi 网页上自动写的提示词是这样的：
写得也还可以用啊。
如果是我自己写，大概率只会写：“参考 3 视图创建 Tripo3D 模型”这样就可以了。 它写得很详细，主要是怕执行模型太弱智。 实际上现在的模型，根本不需要写这么多提示词。
它这个提示词有一点很好，就是规范了输出的内容，指定了生成内容为 GLB 和预览图。
根据生成结果，点击上面的预览之后，就可以点击下载了。
这里有一个难点，就是点击下载的时候会弹出一个另存为，这个时候光调用浏览器就不够了。只能调用整个电脑控制功能才能搞定。
整个过程消耗的时间如下：
从开始到结束，大概就是 20 分钟时间，这个速度不能算特别快，但是也算是比较快的。
token 和费用的情况：
整个过程消耗了 3800 万 tokens，按 API 价格缓存折算以后，大概 56 块钱！
时间和费用大概就是这个样子！ 这个价格你们能接受吗？
它做得不太够好的地方主要在细节。
比如它给网页分派任务的时候，没有选择直接粘贴图片，而是选择弹框选择图片，这样就会麻烦很多。
另外下载生成文件的时候，它是先下载到了 Download 目录，其实它可以直接指定目标路径的，这样就能省掉复制粘贴这一步了。
Kimi Code 的情况就分享到这里了。接下来看看同为国产的 ZCode！
2、ZCode
ZCode 我之前也给大家推荐过很多次，这个工具其实更新也很快，功能也不错了。但是它上次搞出了个大事情...后来就全部开源了！
我先快速总结一下它在这个任务上的情况：ZCode 电脑操作还不是很熟练，时间消耗较多。
然后看一下它的执行过程：
它摸了一下情况后做了如下判断：
到这里基本没有问题，已经清晰找出了 3 个对象。
然后它就开始调用内置浏览器了，我友情提醒它内置浏览器没有登录信息，要调用本地 Chrome 浏览器。
然后它就开始有点呆了：
出现了连续失误，然后反复尝试！
它还出现了“刚才误按了 F5 刷新，无碍。等加载完直接读取对话内容。”！
牛逼，它还能按错 F5！！！
另外有一点很有意思，它直接去抄了一波 Kimi 的提示词。因为 Kimi 官网的记录是有连续性的，ZCode 可以读到之前的对话，所以它就直接抄了一波提示词。
它还出现了一些其他失误：
比如这个地方上传了两张一模一样的参考图。然后它说它要把多余的一张图片删掉，实际上并没有做到，直接写上提示词发送了。
有了制作第一个角色的经验之后，后面的 2 个模型就顺畅了很多。
后来，在生成模型的时候出现了一个异常情况：
就是孙权这个模型，无法生成：
这个问题应该是 Kimi 网页的问题，ZCode 没有找到很好的解决方式！
ZCode 在下载文件的时候，做得也不是太好，一会儿下载 GLB 文件，一会儿下载 Zip 文件，脑子不太清楚的样子。
整个过程大概消耗了 2 个小时，中间停顿，错误尝试太多了，看得我急死了，我真想手动给它点了。
整个过程消耗的 tokens 为……让它查消耗记录它说查不到。
我提示它看 session，它就又懂了，说“你说得对”。
最后的统计数据如下：
整个过程消耗了 2800 万 tokens，转换成 API 价格 7 块钱。
问题和数据都展示完了，说说它的优点。
它的优点是，至少整个流程都完成了，下载的时候直接保存到了目标路径，知道去抄别人的提示词，来保持一致性，还有一点是，价格很便宜。因为我用的是 Flash 模型，没有用 GLM5.3，也有很大的关系。
相比而言，Kimi Code 在这个任务上要清晰干练很多，消耗的时间也只有 1/6。 但是从 token 消耗和费用的角度来说，ZCode 要便宜很多了。
3、Codex
最后，终于轮到 Codex 了。
我也可以先来一个总结：Codex 网页操作非常犀利精准和快速，但是电脑操作出现大问题，配额消耗巨快。
使用 Codex，一开始就遇到了问题。刚开始没多久，就把任务给停了。模型生成成功最后也没有下载到本地。
几乎所有问题都聚焦于： ‘电脑控制工具仍因“无法可靠识别当前浏览器 URL”强制停止’。
我查看了电脑控制功能，全部显示正常！
为了解决这个问题，我单独起了一个对话。让它排查问题：
除了让我更加生气之外，并没有解决这个问题！
它除了道歉之外，就是东扯西扯，完全找不到重点。
刚开始只是机械地重复问题，然后我让它去网上找答案。
它的第一句话，永远是“你说得对”。
第二句话是，终于找到了一模一样的问题。
第三句话是，下载文件依旧失败（其实还是没有解决问题）！
这个问题让我非常失望，我本来准备刨根问底的，但是配额用完了，两个问题加在一起，我就非常生气。我就让它干这么点小事儿，还没干完，配额居然用完了。辣鸡！
不过，它的 Chrome 浏览器操作还是非常犀利的。
整个操作过程思路极其清晰，速度极快。遇到生成模型失败的异常情况之后，也非常机智地解决了。
在生成图片的过程中，它和 ZCode 遇到了一样的问题。就是 Kimi 的在线生成模型功能突然不可用了。提示读取不到图片、或者乱七八糟的问题。
这个时候 ZCode 选择在同一个对话下继续尝试，或者新开一个对话尝试。最终全部失败了！
Codex 很机智地在成功的对话后面补了需求，把两个失败的对话中的内容，直接接在了成功的对话后面。最后成功地生成所有模型。
我后来分析了一下，估计是云端服务资源紧张，新开的对话分配不到，失败的对话自然也分配不到。但是已经成功过的对话，却已经分配到了，可以继续生成。
Codex 里的 GPT-6.1 Sol 似乎是悟出了这一层关系，所以果断地进行了操作。
所以，除了它没法启动“电脑控制”功能之外，浏览器操作是非常犀利的，智商也是在线的。
下面看一下它的数据：
它整个过程只用了 19 分钟，如果不是遇到两个例子失败的情况，它还能快很多。
所以从速度的角度来看，它肯定是最快的。主要是它对各种浏览器操作太熟悉了。
它的 tokens 消耗是 1189 万，也是这几个工具中最少的。
它对应的费用大概是 ¥10.63，比 ZCode 贵了一点点，但是比 Kimi 便宜了好多。
现在所有的工具都展示完了。
整体来说：
KimiCode 电脑操作能力不错，时间比较短，处理比较精准，整个任务可以全程自主完成，就是费用太高了。
ZCode，任务也完成了。但是磕磕碰碰，不太熟练，消耗时间太久了，但是消耗费用最低。
Codex，浏览器操作犀利，模型智商在线，速度也很快，但是出现了一个莫名其妙错误，导致任务频繁中断，最后无法完成全套的电脑操作，最终下载这一步一直无法完成。
我让 Agent 操作的核心需求是全自动完成这种繁琐的操作，Codex 虽然厉害但是并没有做到这一点。
它们一直在宣传电脑操作很厉害，所以我对它是有很高预期的，但是它连电脑控制启动都启动不了，软件里面显示的状态全部正常，但就是用不了，你追问它问题，它也非常敷衍，我提示它去网络搜索，发现了一样的问题，但是最终也没有找到解决方案。
另外一个让我很生气的是它的配额太少了。
配额重置之后，好像又变少了。我今天大概生成了 73 张图片，就直接把五小时配额消耗完了，然后补了 12 张图片，以及测试今天这个问题，问题还没解决五小时额度又到了。
所以就有了标题中说的“气死我咯！让他给 Kimi 提鞋去吧”。Kimi 刚好也是一个配额很少的家伙，你们确实可以凑一对了。Claude 是完全不带你们玩了。
不过话说回来，Codex 做图片还是独一份：
我让它帮我设计了 20 个三国角色，偏亮偏高饱和风格，这三视图做得还是非常稳的。画风也是比较好看的。
让Codex出角色图，让Kimi做成3D模型，这个路径也不错。 我终于挖掘了一个Kimi的用途。
今天就这样吧~~希望通过今天这个对比测试，能给大家一些参考。
另外Codex这个毛病，大家有遇到么？有解决方案么？Codex其实还是很好用的，现在主要是被这个奇怪的问题给卡住了！
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。
I'm finally mad at Codex!
The Invincible Computer Operating Power that this chick has been blowing has collapsed!
And then the problem was left open and the quota was emptied. And then we're gonna merge Codex and ChatGPT quotas.
I used it today, only 48%!
I'm not using Astra, it's Sol.
I clearly feel that the quota is much lower now!
Forget the quota. Tell me about computer manipulation.
This function would have been useful in saving us a great deal of time and energy. All you have to do is give a mission, and it automatically helps you with the computer. It's really cool.
I've just recently had one demand, and I've compared a few Agent tools, and I've found that there's still a big difference.
The main subjects were ZCode (option 5.3 Flash), Codex (option GPT-6.1 Sol) and KimiCode (option K3.1 Pro).
They're all using official Q42s with official models!
Here's my test scenario:
Kimi Network provides a "3D modelling" function that works well. I want them to help me read a set of local reference maps and then automatically call the web, upload them automatically, fill in prompt, download the generated files automatically, and sort the files automatically.
It's a very cumbersome job to do, and it's always waiting for, and it's perfect for Agent to do it.
It took half a day to do it. It's done. It's done. Share information on the operational capacity of the three tools, including time consumption, costs, etc.
1、KimiCode
KimiCode, there's nothing wrong with the whole thing.
Its broad implementation logic is as follows:
I'm going to look at the pictures from Concepts\v2 and confirm which three are the characters and their filenames, then upload them to Kimi, generate, track progress and download the corresponding directory using Tripo3D Multiview.
At the latest generation, the three characters are Sun Tse, Hao Hu, Ganning. I'll use the three-view version of the Views to save the model to pubg\docs\Tripo3D.glb\v2, corresponding to the existing V1 model directory.
Reference diagram confirmed and archived. I'm going to create a new three-country V2 Tripo3D group in Chrome to create a task-by-tape group; these tabs will be kept for tracking and you can shut me down at any time.
The extended direct upload of permission has not been opened, and I have switched to computer control to open the Windows file selection for upload, without you changing the browser settings.
Sun Tsai 3 view uploaded and prompt completed. The Tripo3D generation task is now submitted and normal access to the mapping and modelling process is observed.
Sun Tsai has entered the triview split and Tripo3D generation process. In order to reduce the waiting, I will continue to submit to you on the independent tab page for a single follow-up of the three tasks.
All three of the characters have been successfully commissioned and are being generated backstage. I'll check the schedule on a rotational basis and download it to V2, as soon as it's done.
Sun Tsai has been produced and GLB contains a 4K poster about 51 MB. I'll download Sun's policy first, and the other two continue to produce.
The idea of the whole process is clear!
The only problem was to upload permission, and immediately to upload it with the Windows File Selector!
It automatically wrote prompt on Kimi's website:
It works.
If I write it myself, I'll probably just write, "Refer to 3 View to Create Tripo3D Models." That's it. It's very detailed, mainly because it's too retarded to implement the model. Actually, there's no need to write so many prompts.
It's a good prompt that regulates the content of the output, assigns the content to GLB and preview.
Based on the results, click on the preview above and then click to download.
One of the difficulties here is that when you click on the download, you will pop up a remnant. This can only be done by calling the entire computer control.
The duration of the process is as follows:
From the beginning to the end, it's about 20 minutes, and it can't be that fast, but it's also pretty fast.
Token and costs:
The whole process consumes 38 million tokens, converted at API cache, about 56 dollars!
That's probably the time and cost.
Where it is not doing well, it is mainly in detail.
For example, when it assigns tasks to a web page, it does not choose to paste the picture directly, but instead chooses a frame to select the picture, which is a lot of trouble.
When you download the generated file, it downloads to the Download directory, which in fact directly specifies the target path, thus saving the copying of the paste.
Kimi Code was shared here. Now look at ZCode!
2、ZCode
Zcode, I've recommended it to you many times before, and it's a tool that's very quickly updated and functional. But it made a big deal last time... and then it all started!
Let me get this straight: Zcode's not very skilled and time consuming.
Then look at its implementation:
It touched the situation and made the following judgement:
There's almost no problem here. Three objects have been clearly identified.
And then it started calling the built-in browser, and my friendship reminds it that the built-in browser has no log-in information, and to call the local Shrome browser.
And then it started to get a little bit silly:
There's been a series of mistakes and repeated attempts!
It also appears, "Just wrongly pressed F5, refreshed. When the dialogue is loaded and read directly.”
F5!
And one interesting thing, it's just copying a wave of Kimi's prompt. Because Kimi's records are continuous, and Zcode can read the previous conversation, it just copied a wave of prompt.
It also had some other failures:
For example, this place has uploaded two identical reference maps. It then said that it was going to delete a surplus picture, which was not actually done, and that it was sent directly on prompt.
With the experience of producing the first role, the two models in the back were much smoother.
Later, there was an anomaly in the generation of models:
It's Sun Kwon.
It's supposed to be Kimi's web site, Zcode didn't find a good solution!
Zcode didn't do too well when he was downloading the file, downloading the GLB file, downloading the Zip file, and not thinking clearly.
The whole process took about two hours, stopped in the middle, tried too many mistakes, and I was dying to give it a little hand.
The whole process consumes tokens to make it look at the consumption record and it says it can't.
I hinted at the session, and it understood, and said, "You're right."
The final statistics are as follows:
The whole process consumed 28 million tokens, converted to API price of $7.
Both the problems and the data have been shown and their merits have been said.
It has the advantage that at least the whole process is complete, that it is downloaded directly to the target path, that it knows to copy another person ' s prompt for consistency, and that it is cheap. Because I'm using the Flash model, and I'm not using GM5.3, and I have a big connection.
Compared to Kimi Code, the task is much more clearly capable and consumes only one sixth of the time. But from the point of view of token consumption and cost, Zcode is much cheaper.
3、Codex
Finally, it's Cordex's turn.
I could start with a summary: Codex is very precise and fast, but there's a big problem with computer operations, and quotas are running fast.
Using Codex, problems were encountered at the outset. We stopped the mission very soon after. The model was successfully generated without being downloaded locally.
Almost all of the problems are focused on: `The computer control tool is still being forced to stop because the current browser URL cannot be reliably identified'.
I checked the computer control. It's all normal!
In order to solve this problem, I initiated a separate dialogue. Let it sort out the problem:
It didn't solve the problem except to make me more angry!
Besides apologizing, it's bullshit. It can't get any focus.
It's just a mechanical repetition of the question, and I'll let it go online.
Its first sentence will always be “You're right”.
The second sentence is that the same problem has finally been found.
The third sentence is that the download still fails!
I was very disappointed by this issue, which I was prepared to take to the bottom, but I was very angry when the quotas were run out and the two questions combined. I'm just letting it do a little bit of work. I'm not done yet. Chicken!
However, its Chrome browser operation is very sharp.
The whole operation is very clear and fast. Anomalous circumstances that had failed to generate the model had also been very cleverly resolved.
In the process of generating pictures, it had the same problems as Zcode. It's that Kimi's online generation model is suddenly useless. It's not easy to read pictures, it's a mess.
This is when Zcode chooses to continue to try under the same dialogue or to try a new one. It all failed!
Codex was resourceful in filling the needs behind a successful dialogue and linking the two failed dialogues directly to a successful dialogue. Eventually all models were successfully generated.
I later analysed that it is estimated that cloud-based services are stretched, that new dialogues are not distributed, and, naturally, failed ones are not. But the dialogue that has been successful has been allocated and can continue to be generated.
GPT-61 Sol in Codex seems to have learned about this relationship, so it's done it decisively.
So, in addition to its inability to activate computer control, the browser operation is very sharp and the IQ is online.
Look at its data:
It took only 19 minutes for the whole process, and it would have been a lot faster had it not been for two cases of failure.
So from a speed perspective, it must be the fastest. It's mostly too familiar with the browser operations.
Its tokens consumes 1,18.9 million, the least of these tools.
It's about 10.63, a little bit more expensive than ZCode, but much cheaper than Kimi.
Now all the tools are out.
Overall:
KimiCode has a good operating capability, a shorter time frame, a more precise processing process, and the entire task can be done autonomously, which is too expensive.
ZCode, the mission was completed. However, it is less skilled and consumes too much time, but at the lowest cost.
Codex, the browser operation is sharp, model IQ is online and fast, but there is an inexplicable error that leads to frequent interruptions and ultimately to complete the full set of computer operations, and the final download has been impossible.
The core need for Agent to operate is to do it all automatically, and Codex, though powerful, did not.
They've been promoting computer operations, so I have high expectations, but they can't even start the computer, they're in a normal state, but they just can't, and you ask questions, and it's very perfunctory, and I'm telling it to search the network and find the same problem, but eventually they don't find a solution.
Another thing that pisses me off is that it has too few quotas.
After the quota had been reset, it seemed that there had been less. I've got about 73 pictures today, and I've run out of five-hour quotas, and I've filled in 12 pictures, and I've tested the problem today.
So the title says, "Fuck me! Let him take Kimi's shoes." Kimi happens to be a guy with very few quotas, and you can really make a couple. Claude is totally out of your league.
But then again, Codex is still the only photo:
I made it work for me to design 20 three-country characters, to highlight high saturation styles, which are still very steady. It's better to paint.
Let Codex make a role map and Kimi make a 3D model, which is a good path. I finally dug up a Kimi use.
Let's do it today. I hope this comparison test will be useful.
And the codex problem. Has anyone met? Any solutions? Codex is really good at it. It's mainly stuck to this strange problem!
Auto-fetched and translated by AI Atlas. All rights belong to the original author.