Five LLMs Tried To Build A Website. ChatGPT Failed. The Model That Shipped Was The Biggest Surprise.
Real builds vs pretty prototypes across Claude, Kimi K2, Gemini, MiniMax and ChatGPT.
Large language models are exceptionally good for website building. If I were to pick only one job that will see a significant reduction in workforce, I would confidently say frontend developers. And if I were to pick one area of IT solutions that will become obsolete, it is low code. In my opinion, low code, where you click together an interface like in Wix, has no future at all. In this article, we will show you why.
Olga Chatelain and I built five websites in a day based on her CV. How this happened to be, how faithful websites are to the CV, how long it took us is all in her article. You should read it to have the full picture.
We need your help to vote on which website is the best. The one you vote for will become her official website, so choose responsibly!
RULES:
if you like the website - vote for it
if you do not - go back to “see all websites” and check other implementations
you can vote for multiple websites
Click here to review the websites.
Based on the votes, I will pick two LLMs for the course on 16 December on how to build a website like this, and we will build two websites together in only 90 minutes and deploy them. All paid subscribers are invited to the session. Even if you cannot join live, you will receive all the prompts, detailed step by step instructions, the Git repository, slides, and the course video. More information here:
This course is also part of an AI Realist offering for companies, marketing agencies, and others who want to upskill their teams. If you want a private course where we build a website for your business, deploy it to production, and teach your team how to maintain it, contact me at info@mariasukhareva.com.
Executive summary of the results
MiniMax (China, open-weight): The only one that acted like a real web developer. It built a full site that you can actually put online. A bit heavy, but it works end to end.
Claude (US, proprietary): Made a beautiful single-page site. Looks polished and reads well, but it is missing forms and other pieces you need for real use.
Kimi K2 (China, open-weight): Tried to do a lot at once. Multiple pages, quizzes, charts. Ambitious, but too many parts felt unfinished, so it breaks easily.
Gemini (US, proprietary): Great visual mockup. Modern look, responsive, nice icons. Still more of a picture of a site than a working site.
ChatGPT (US, proprietary): Clean, simple page that is easy to host. Almost no interactive bits, so it feels shallow.
We also open-source the websites’ implementations for your own validation
Bottom line: MiniMax gives you something you can ship. Claude and Kimi K2 are promising drafts that need more work. Gemini and ChatGPT are good starting designs, not finished websites.
Experimental setup
Goal: build a professional website for a highly qualified AI and blockchain expert and coach Olga Chatelain who offers her services as a freelancer. For the website implementation, we used her CV and the color scheme from Substack Blockchain meets AI. You can read more about the detailed setup in Olga’s post that complements this one. It covers the quality of the generated websites, how faithful they were to the CV, and some of the model-related limitations.
If you like this newsletter, consider upgrading your subscription or visiting the shop
Check out our website: https://www.airealist.org/
Paid subscribers get:
Priority answers to your messages within 48-hours
Free participation in online training “How to build a website with AI“ on 16th of December
Founding members receive even more:
A 45-minute one-on-one call with me
High-priority personal chat where I quickly reply to your questions within 24-hour
Support independent research and AI opinions that don’t follow the hype.
— or check out the anti-hype shop
Implementation Breakdown By The Model
Claude Implementation
We used a Claude Pro subscription that costs $15 a month. It includes access to Opus 4.5 and Claude Sonnet 4.5. We used Opus 4.5 as the best possible model. We did not use Claude Code, as this tutorial is for non‑coders and we worked solely with the chat interface.
Claude Opus 4.5 did very well with building a simple HTML‑based website. Its approach was very straightforward in terms of structure – everything in one file including CSS – but the implementation does include a small, fairly sophisticated JavaScript section (e.g. an IntersectionObserver‑based scroll animation and smooth scrolling).
There were major limitations. Claude does not generate visuals, so for images we had to use MiniMax deployment of Nano Banana Pro. Another limitation was discovered right after we uploaded the images – the Claude preview does not show them, so to see whether the website actually looked good with the images, we had to move to VS Code and use a local deployment of the website for preview. However, as it is just one file, we used VS Code solely for previewing and continued generating the code in the same chat, copy‑pasting it into VS Code.
Then we hit the next limitation – we ran out of Claude Opus 4.5 quota. We bought the subscription just for this experiment, so we were not familiar with how Claude works and assumed we could just switch that same chat to a cheaper model – but you cannot change the model for an existing chat. The website was only halfway finished, so we had to switch to VS Code, where I have a GitHub Copilot (Pro+) subscription, and continue there.
While Claude did not need many major adjustments for the website and understood fairly quickly what was needed, it is still relatively expensive to use Opus 4.5, and the need to generate images elsewhere and not being able to preview them directly is somewhat of a nuisance.
Coders can review the implementation here: https://github.com/ktoetotam/website-building-blockchainwithAI/tree/main/claude
You can see the resulting website here: https://ktoetotam.github.io/website-building-blockchainwithAI/claude/olga-chatelain-portfolio.html
If you like the website click on vote, if you do not like it, click “See All“ and check other implementations.
Gemini Implementation
We had very high expectations for Gemini. Just as with Claude, We bought a yearly subscription to Gemini Advanced for $21.99/month.
We chose the Gemini 3.0 Pro model and gave it the same prompt as all the other models as the starter. It generated the HTML, and at that point the fun with Gemini Advanced was over, because it does not have a built-in website previewer.
It is so far the only system that does not provide one, so I had to package the generated code and move it to Visual Studio Code with my GitHub Pro+ subscription. I used the Gemini 3.0 Pro Preview in GitHub Copilot to continue working on the Gemini version.
Similarly to Claude, I used Nano Banana 3.0 in Minimax and added images manually. Gemini has generated a very simple, long HTML file and put everything into one file. The result was good, easy to create, and stable, but it required external tools like VS Code.
Correction: Gemini does have a Canvas preview where you can see the website. However, as I have currently tested it, it is very slow and it tends to substantially change the whole website each time a small change is asked. It allows preview of the generated images. VS Code is a much more pleasant way to build a website with it.
Coders can review the implementation here: https://github.com/ktoetotam/website-building-blockchainwithAI/tree/main/gemini
You can see the resulting website here: https://ktoetotam.github.io/website-building-blockchainwithAI/gemini/gemini-index.html
If you like the website click on vote, if you do not like it, click “See All“ and check other implementations.
MiniMax Agent Implementation
This one turned out to be the biggest surprise for us. We did not expect it to deliver a very competitive website, but man, were we wrong.
We used the MiniMax Agent and started with the same prompt as always. MiniMax took a very different approach from all the other LLMs. It started building a complex website using TypeScript, JavaScript, and React-based frameworks. It also generated all the images and placed them directly on the webpage.
It was fascinating to watch how MiniMax creates websites: it takes a screenshot of the generated site, analyzes the design, and then applies changes. It also tests that all the links work and that everything is clickable.
The obvious drawback is that IT TAKES FOREVER. One prompt can keep MiniMax busy for anywhere from 5 minutes to an hour. The longer the chats go on, the more MiniMax seems to overthink. At some point I asked it to review the design and make it a bit prettier, and the “overthinker” kept going for about 2 hours. It produced an output, but then I wanted one small change, which sent it into another hour of rambling and in the end it could not deploy the solution. I stopped it and just changed the background manually.
This was the very first version of the website and you can explore it here as MiniMax deployed all the versions:
https://eelhpxr1onox.space.minimax.io/
An interesting observation was that MiniMax had a strong tendency to hallucinate certain sections, such as “Client Success Stories,” and it was surprisingly difficult to get rid of them. Olga writes in her article:
The most amusing hallucination came from MiniMax. The model desperately wanted my website to include a section called “Client Success Stories.” Every second prompt included it. We deleted it. It reappeared. We deleted it again. It came back. Like a very confident ghost that simply believed this section belongs to me.
MiniMax Agent with 5000 credits a month costs 17 euros. Since the credits generously provided by MiniMax some time ago had already run out, I had to top it up for 17 euros but building the website took around 3000 credits. This is a very competitive price considering that all images were generated there.
Another interesting fact about MiniMax is that M2 model is open-weight and can be deployed on your infrastructure.
All in all, MiniMax Agent is the most well-rounded chat LLM for building a website in its own environment. It supports image integration, JavaScript and TypeScript generation, deploys the website, and lets you test it. The downside is that it can take forever, and the longer the chat gets, the longer it thinks. In my opinion, it is very well suited for prototyping more complex websites, but eventually you would want to move the project into VS Code or another IDE and continue coding there.
Coders can review the implementation here: https://github.com/ktoetotam/website-building-blockchainwithAI/tree/main/minimax
You can see the resulting website here: https://ktoetotam.github.io/website-building-blockchainwithAI/minimax/olga-chatelain-website/dist/index.html
If you like the website click on vote, if you do not like it, click “See All“ and check other implementations.
Kimi K2 Implementation
This was the model I was most excited about, because I expected a strong agentic model to produce a very capable website. Surprisingly, it was not as impressive as I had hoped. Olga was responsible for generating the website with Kimi K2 (you can read more on her blog), and it took her many reruns to get to a reasonably good version.
Similarly to MiniMax, Kimi K2 generated a more complex website with JavaScript, a modular multi‑page structure and a separate resources folder. There are four HTML pages (index, services, about, insights) and three JavaScript files (main.js, services.js, insights.js) powering a service configurator, scroll animations, and an (incomplete) ECharts‑based insights page. The setup is not as heavy as MiniMax (no React, TypeScript, or build system), but still significantly more interesting and complex than the Western models, which mostly produced a single HTML file with minimal JavaScript.
The first version of the website was somewhat buggy and clearly unfinished. A big plus of Kimi K2 is that it also generated images and showed their previews directly in the chat. The final project includes 12 AI‑generated PNGs (for articles, consulting, hero section, etc.). However, the image model still struggles with spelling inside images, so it is safer to generate visuals without text or to overlay your own text afterwards.
Just like MiniMax, Kimi deploys the website on their own servers. Another cool feature that appears unique among all the tools I tested is a built‑in preview of the mobile version. The generated code and design docs also follow a mobile‑first approach with Tailwind responsive classes and explicit notes about simplifying animations on smaller screens.
Coders can review the implementation here: https://github.com/ktoetotam/website-building-blockchainwithAI/tree/main/kimi-k2
You can see the resulting website here: https://ktoetotam.github.io/website-building-blockchainwithAI/kimi-k2/index.html
If you like the website click on vote, if you do not like it, click “See All“ and check other implementations.
ChatGPT Implementation
Olga Chatelain was also responsible for building the website with ChatGPT, and this was probably our biggest surprise after MiniMax – in a bad way. The experience was not great. ChatGPT kept hanging during generation, and although it can theoretically produce images, it didn’t render any image previews in the UI. As a frontend for website building it was barely usable.
The final code reflects this: ChatGPT produced a single, very long index.html file (about 1,150 lines) with embedded CSS and almost no JavaScript – just one line to update the year in the footer, plus shared vote-widget.js which I built. There is no real interactivity, no modular structure, and no build system or type safety. The design itself is clean and accessible, with a well‑structured CSS variable system and decent responsiveness, but the implementation is essentially a static one‑pager.
And the result… you can judge for yourself. To keep the comparison fair and visually closer to the other sites, we manually added the hero banner image afterwards by editing the HTML, since ChatGPT never managed to generate and wire an image asset into the page on its own.
Cost of Implementation
Kimi K2 and MiniMax Agent are the most affordable options in this test. MiniMax Agent costs 17 euros for 5000 credits, which is enough to build one or two full websites (we spent around 3600 credits). Kimi K2 uses a similar credit-based system and stays in a comparable price range; however, the 11 euro subscription was not enough to polish the website, so the version you see is unpolished.
Claude Pro is noticeably more expensive, and its daily limit for Opus 4.5 was not enough for full website development. It ran out of quota long before the site was finished, so you need to switch to cheaper models mid build.
ChatGPT was hard to evaluate because it kept hanging on a MacBook Air M4 1 TB, and the website quality was disappointing. You probably would not run out of credits, rather of patience.
For the remaining work, we used VS Code with GitHub Copilot Pro+. Continuing the project with Opus and Gemini Pro through Copilot consumed only about 5 percent of the GitHub credits, which made it the cheapest way to finish the sites once the LLM quotas were exceeded.
We are encouraging you to review the websites and vote for your favourite. We will share the votes in a couple of weeks. You are also very welcome to review the code in the git repository and give us feedback if you agree with our evaluation. Also, join the course on the 16 December to build a website together.
Really nice piece od applied researxh there. And a fun read
I liked the sarcasm :)