Which AI Can Create the Coolest Web Page? GPT 5.2 vs Gemini 3.0 Pro vs Opus 4.5 vs bolt.new vs v0 vs Lovable
Introduction
As the first step in my smartphone app development journey, I am developing "Sumineko", a very simple smartphone app that lets you watch random cat videos endlessly in the corner of your screen while multitasking.
Sumineko (Under Development)
To register this app on official stores, I need to prepare a privacy policy page on the web, so I decided to take this opportunity to create a Landing Page (LP) for the app.
Therefore, in this article, I will have representative LLMs and AI coding services create the foundation of the LP, and I will compare their results based on my own prejudice and bias.
Contenders
| Contender | Features |
| v0 (v0 Agent) | Vercel's agent tool. Supports everything from UI implementation to debugging based on prompts. |
| bolt.new (Claude Sonnet 4.5) | Browser-based Claude agent. Efficiently builds Next.js apps in a conversational format. |
| Lovable | A builder supporting multiple AI models. React+Supabase apps can be generated and edited intuitively. |
| Gemini 3.0 Pro | Google's latest high-performance model. Quickly generates sophisticated Web pages with multimodal input. |
| Opus 4.5 | Anthropic's strongest model. Leverages long context to carefully handle complex Web app design. |
| GPT 5.2 | OpenAI's latest model. Compared to GPT-5.1, it has 1.8x the "correct answer rate," significantly improved "code accuracy," and 30% fewer hallucinations. |
WARNING
Note on Treating Them as Equals
Pure LLMs (Gemini 3.0 Pro, Opus 4.5, GPT 5.2) and development support systems with built-in LLMs (bolt.new, v0, Lovable) have different system architectures. However, for the purpose of this article, I am treating them as equals for convenience, aiming to subjectively compare the quality of web page generation results using the same prompt. Please note that this experiment lacks strict rigor.
Prompts
I prepared the following two patterns to loosely verify their ability to ad-lib and their faithful adherence to requirements.
Minimal Pattern: States only the minimum requirements (Hands-off).
Detailed Instruction Pattern: Describes relatively detailed requirements.
Minimal Pattern
A pattern where I primarily provide a simple overview of "Sumineko" and the necessary sections, leaving the design and wording up to the AI.
Detailed Instruction Pattern
A micromanagement pattern that, in addition to the overview and necessary sections, gives detailed instructions on the structure of each section, wording, and design aspects like color schemes.
Results Summary
NOTE
Although they are responsive, this article only deals with the desktop view.
Minimal Pattern
v0
bolt.new
Lovable
Gemini 3.0 Pro
Opus 4.5
GPT 5.2
Detailed Instruction Pattern
v0
bolt.new
Lovable
Gemini 3.0 Pro
Opus 4.5
GPT 5.2
Looking at the Results Individually in Detail
v0 (v0 Agent)
Minimal Prompt
The Minimal Pattern has a super standard UI design with a very reserved color scheme, giving the impression of being "safe" in both good and bad ways.
Since the images, videos, and icons are placeholders, the simplicity might be standing out even more.
Perhaps if actual assets were inserted, this simplicity would provide a good balance that highlights elements like screenshots and videos.
Detailed Instruction Prompt
On the other hand, the Detailed Instruction Pattern faithfully follows the prompt's content, but generally results in a similarly reserved design.
It doesn't use bold text for emphasis, giving a rather flat impression—for better or worse—not much different from the Minimal Pattern.
bolt.new (Claude Sonnet 4.5)
Minimal Prompt
The Minimal Pattern features a soft, lively design and color scheme based on orange. The eye is naturally drawn to the orange text and images, giving a sophisticated impression.
Appeal text visible without scrolling
It's a small point, but having a subtle appeal element like "Experience the magic" in the HERO section—meaning it's visible without scrolling—is a good touch to reduce bounce rates.
However, even though they are placeholders, the design of the smartphone in the HERO section and the flatness of the elements in the collection section felt somewhat dated, cheap, or amateurish.
Detailed Instruction Prompt
Conversely, the Detailed Instruction Pattern is very faithful to the prompt, which is good, but it gives the impression that this faithfulness killed bolt's strengths.
It's definitely not bad overall, but compared to the Minimal Pattern, it's too monotonous.
This isn't bolt's fault; rather, it's likely that the low quality of the prompt I prepared is reflected directly in the form.
Lovable
Minimal Prompt
For the Minimal Pattern, the first thing that catches the eye is the illustration of a cat with a collapsed face.
Collapsed face cat 1
Collapsed face cat 2
Since I didn't provide assets, it seems to have kindly prepared them for me, but the surreal design looks like it was drawn by a kindergartner, which makes me chuckle.
However, these are just temporary placeholders, so if you ignore the collapsed-face cats and look at the design of the other parts, it's not bad—there are cats scattered in the background, and parts to be emphasized are properly highlighted with brown or orange accent colors.
Detailed Instruction Prompt
The Detailed Instruction Pattern is very faithful to instructions, but it also shows ingenuity throughout, such as using bold text for strings I wanted to emphasize, using a soft font to match the app, and highlighting the PiP transition button in orange, which wasn't explicitly specified in the prompt.
Soft font matching the app
Button that stands out properly against background contrast
Gemini 3.0 Pro
Minimal Prompt
The Minimal Pattern lacks playfulness—for better or worse—and features an overall tight design/color scheme, giving the impression of a business app or tool rather than a healing app.
It's just my bias, but the direction feels very "Google-like."
Regarding the design direction, it clearly doesn't match the app, but putting that aside for a moment, I feel the overall quality is extremely high.
It gives the impression of "I made this following the Web Design Textbook!" It has no quirks, and I feel I could trust it with a request.
Representation of PiP
Representation of Browser
It's a small detail, but the reproduction quality of the placeholder illustrations for PiP and the browser is high.
It makes it easy to visualize the finished product, which is a surprisingly nice point.
Detailed Instruction Prompt
The Detailed Instruction Pattern is faithful to instructions like the other AIs, and while it is dragged down by the low level of design instruction in the prompt, it ad-libbed a dynamic design where the smartphone image and collection screen image are slightly tilted—a small detail, but a very good accent.
Tilted Smartphone
Tilted Collection Image
Also, the cat casually walking in the footer is stylish.
Cat walking on the footer
Opus 4.5
Minimal Prompt
For the Minimal Pattern, the design is generally bright, colorful, and cute, clearly reflecting the theme of "Cat Healing App."
The header also has a transparent, modern design that is very stylish.
Colorful and unified design
With icons scattered in the background, gradients in headings, and a colorful palette, it seems to be an excellent design that is consistent and doesn't feel disjointed, while still being full of playfulness.
Detailed Instruction Prompt
On the other hand, the Detailed Instruction Pattern completely killed the good qualities of Opus 4.5 when compared to the Minimal Pattern, resulting in a suddenly boring design.
Like bolt, it appears the low level of the prompt was directly reflected in the form.
GPT 5.2
Minimal Prompt
The Minimal Pattern has a design and color direction similar to Opus, using colorful gradients as accents throughout for a bright atmosphere.
When an amateur like me tries to use such varied gradients, it inevitably looks cheap, but here it gives the impression of being put together beautifully and skillfully.
While other models expressed placeholder assets with cat emojis or photos, GPT 5.2 seems not to use such things at all.
Also, differences from other models are prominent in the HERO section.
HERO Section
The entire HERO section is represented as a card element, and the smartphone image is represented as a square rather than a vertical rectangle.
However, elements at the bottom of the smartphone screen are protruding, and the line with download links is placed on the same level as the app's appeal text, making it cluttered. Even as a design amateur, I feel a slight sense of incongruity.
Detailed Instruction Prompt
The Detailed Instruction Pattern became a boring design, similar to Opus and bolt.
As I've mentioned many times, I think this is a manifestation of being faithful to the prompt, for better or worse.
However, while Lovable added orange as an accent color and used a softer font to match the app, and Gemini intentionally tilted elements, showing sparks of expressive ingenuity, absolutely nothing of that sort was seen in GPT 5.2.
General Review and Conclusion Dripping with Personal Bias
My Favorite AI Ranking
Based on the results obtained this time, I will rearrange the AIs in the order I liked, along with my reasons.
By the way, I am a complete amateur in terms of design and have no professional knowledge, so this evaluation is strictly based on an amateur's intuition and experience.
| # | Contender | Reason |
| 1st | Gemini 3.0 Pro | Whether instructed loosely or in detail, I felt a sense of design sophistication and stability in both cases. I felt most strongly that if I just ask Gemini for now, it will churn out a result that is comprehensively above average, so it's 1st place. |
| 2nd | Opus 4.5 | In the Minimal Prompt, it interpreted the app's intent well and reflected it in the design with high quality, which is wonderful. However, I felt caution is needed because if the prompt quality is low, the output quality drops sharply compared to Gemini. |
| 3rd | GPT 5.2 | The overall quality is high and I feel great potential, but I did feel a touch of anxiety about the design of the HERO section in the Minimal Pattern. |
| 4th | bolt.new (Sonnet 4.5) | In addition to the cheap feel of some elements in the Minimal Pattern, it's similar to Opus 4.5 in having high ad-lib ability, but I felt it requires caution as the prompt might kill bolt's good points, so 4th place. |
| 4th | Lovable | As expected, the collapsed-face cats are distracting. Since they are only temporary placeholders, maybe I shouldn't mind, but if the placeholder images are too sloppy, it's hard to visualize the finished product, so I personally have to deduct points. However, the Detailed Instruction Prompt had a level of perfection that rivals other AIs. |
| 5th | v0 (v0 Agent) | Overall safe, plain, and boring. Ad-lib ability is low, and I felt it was strongly dragged down by the user's low level of prompt engineering, so it seemed to require considerable caution for those with strong design preferences. Since it was a tool I used to love temporarily in the past, it's a bit disappointing. |
It's been a hot topic on social media recently that Gemini 3.0 Pro can make anything, but I take my hat off to its excellence in design as well.
I have mainly used Claude Code (Max Plan) and Cursor until now, but this result strongly made me think that switching to Antigravity might be an option.
People Without Design Sense Should Not Overstep with Prompts
I think everyone felt this looking at the results, but in the Detailed Instruction Pattern, while some models showed ingenuity, the outputs basically became plain and boring.
In particular, specifying only brownish colors to match the app's theme colors seemed to spur on the most monotonous designs.
Since AI is basically faithful to instructions—for better or worse—this may be natural, but people without design sense should not stretch themselves to micromanage the design parts in their prompt engineering.
By leaving areas you don't understand well to the AI to some extent, the AI should make judgments based on best practices in that domain.
Doing so inevitably brings out a "I've seen this somewhere before" feeling, but I felt it was far better than an amateur messing around randomly.
The Real Hell Starts Here
The deliverables this time are merely starting points; based on these, a process of repeatedly aligning directions with the AI to get closer to one's ideal UI/UX is necessary.
In my case, despite having no sense, I am strangely particular about UI, so this process takes especially long.
Considering that, factors like the shortness of the AI's thinking time (shortness of trial-and-error spans), low bug rate, and the ability to accurately grasp the intent of instructions and connect them to implementation are actually far more important than the first deliverable.
From that perspective, the ranking order would likely be completely different from the one listed above.
Therefore, it is definitely premature to judge superiority based solely on these results, so I hope you will try them out yourself as a reference.
