Real Project Evaluation: Kimi3 vs Opus4.8 and Fable5

I initially ran some small tasks on Kimi3. My immediate impression was that its output quality was decent, yet it worked very slowly. I’ve seen numerous case studies claiming Kimi3 excels at frontend development, outperforming Fable5 in certain scenarios. GPT remains my primary model, though it struggles with frontend work. That made me curious about how capable Kimi3 truly is and whether it can compete head-to-head with Fable5. To cut to the chase: it is indeed powerful. When pricing is taken into account, it serves as a cost-effective alternative to Fable5. Below is the testing process:
These are the current color schemes for my website. They look rather plain, and my aesthetic sense isn’t great either. I can’t figure out how to optimize them, so I’ve left them untouched all this time.

Prompt for this task:

The color scheme of the TokenDos website does not look appealing. Add a theme setting with reference to the screenshots, but do not copy them entirely. My goal is to make the overall website style match TokenDos's business, while looking elegant, attractive to users.

Here is the reference screenshot

Although this only involves front-end adjustments, there are several hidden pitfalls in the requirements that I did not explicitly point out. I want to see if the model can identify them on its own.

I used the OpenCodeGo plan for Kimi3. The entire task took roughly 25 minutes, which exhausted my weekly quota in one go. The process got stuck at the compilation and testing stage. Fortunately, all the code had been written, so I stopped the task immediately to see if it could work on the first try.

Performance of Kimi3
The project initialization went smoothly. Recently released models rarely make code syntax errors. Existing issues mostly stem from misunderstanding requirements and missing implementations. Let's check the raw output from Kimi:

My first impression was, hmm, it was just alright and fairly standard. I provided reference images, so I secretly hoped it could generate a style more appealing than the references. However, Kimi faithfully reproduced the effects shown in the reference pictures. One point worth praising is that nearly all required modifications were completed. Pages functioned properly with no bugs when switching styles. The login and registration pages were overlooked, which is not a major issue and can be fixed in another round of conversation.
There is a minor problem with the text color on the user management page. In addition, the fund management page was not fully revised and still retained the original color scheme.

❖Performance of Opus4.8
Now it’s Opus4.8’s turn. It was the benchmark just two months ago, yet now it feels outdated. I can’t help but find fault with it. Its working duration is somewhat short, and it also comes across as overly confident.

This is the raw output effect:

It went wrong. First off, it modified pages that shouldn't have been touched. This is the pitfall I was talking about. I only wanted changes for the TokenDos module. It failed to notice this without explicit prompts, and the revisions were incomplete as well.

Second, changes that should have been made were omitted; the entire admin backend was left untouched.

Thirdly, there is a bug in color switching. Only the light mode works when switching to the latter several color schemes, while dark mode functions abnormally.

❖Performance of Fable5
Now comes the final heavyweight, Fable5. The whole process took around 35 minutes. Kimi is already fairly pricey, yet Fable5 burned through my budget painfully. What’s more, the workflow got stuck in the debugging phase, repeatedly running tests to check for code issues. I couldn’t bear the ongoing cost any longer, so I told it firmly to stop testing and that I would conduct the tests myself.

It also adheres closely to the reference images. It didn't bring me many pleasant surprises, yet the output is very comprehensive.

There is also a minor issue with the text color on the user management page. Besides, the fund management page hasn't been fully revised, and the login and registration pages have also been omitted.

Comparison with Kimi3: Both deliver highly complete results and get the task right on the first run. Fable takes more minor details into consideration than Kimi and asked me in advance which color scheme I preferred as the default. Kimi provided two extra color options. Most pages look extremely similar, even sharing the same errors. Kimi features a more concise reasoning process, and its total token consumption is only around one-third of Fable’s. If cost is a priority, I would pick Kimi.

But here’s the catch: Kimi dealt me a heavy blow and left me no choice — it’s no longer available for purchase. Anyone want to team up to get access?