I have tested enough AI coding tools to know that building a website is the easy part. The real challenge is making it feel intentional. A basic layout, a few sections, and some placeholder animations are no longer enough, especially when tools like Claude Code, Codex, and Google Antigravity all claim they can handle serious development work.
So, I gave all three the same complex website brief and watched how they approached it from start to finish. By the end of the test, one tool didn’t just generate code. It felt like it understood what a finished website should actually be.
A word on the prompt and selected models for the test
For this comparison, I didn’t want to give these tools an easy landing page and call it a day. The prompt was deliberately demanding: a graphic-heavy sneaker launch website with bold visuals, interactive colorways, a countdown timer, product storytelling, specs, a comparison section, a gallery, and polished mobile behavior.
In short, it was complex enough to expose weak design judgment and lazy execution. I also brought out the heavyweights for this test. I used Gemini 3.5 (High) in Google Antigravity, GPT-5.5 (High) in Codex, and Fable 5 in Claude Code, so each tool had its best shot at the task.
Codex
Impressive UI, but not quite there yet
Codex was easily the fastest tool in this comparison and quickly showed it understood the overall direction of the prompt. It captured the intended aesthetic with a sleek, gradient-heavy background, sharp geometric elements, and a futuristic feel that aligned well with the sneaker launch concept.
It also followed the checklist closely, including features like the colorway selector, countdown timer, specs, product sections, and FAQ.
However, as I spent more time reviewing the site (and comparing it with others), some issues became more apparent. The typography stood out as the biggest weakness. Codex relied on very similar font styles throughout, which made it difficult to distinguish between headers, subheaders, and body text.
The oversized headings across multiple sections gave the page a louder, less refined tone than expected.
The product visuals were another drawback. For a graphic-heavy sneaker site, the shoe image should be a focal point, but the generated design didn’t convincingly resemble a real sneaker.
The ‘Material Anatomy’ section also felt underdeveloped, with random numbers that lacked clear purpose or visual clarity.
Overall, Codex delivered a functional and complete site, but the execution lacked the polish, balance, and design maturity needed to feel like senior-level work.
Google Antigravity
Came close to taking the top spot
Google Antigravity was a close second in this comparison, and there was a lot to like here. The generated images, typography spacing, and overall presentation felt top-notch.
It understood the graphic-heavy sneaker brief much better than Codex and created a website that looked futuristic, sharp, and performance-focused. The entire design had a sci-fi edge, and it still felt aimed at a younger audience that would actually care about a limited-edition sneaker drop.
One of my favorite parts was the engineering breakdown section. Instead of placing random labels around the shoe, Antigravity made the section feel interactive and intentional.
I could hover over each number and learn more about the material structure, grip, and other design details. That small touch immediately made the website feel more premium and thoughtful.
The copy was also strong throughout, with each section neatly designed and written to match the product’s high-performance identity. However, the hero section held it back from being the winner.
Since the hero is the first thing everyone notices, I expected something more dramatic. Compared to Claude Code’s output, it felt a little restrained. A larger shoe visual, stronger gradient elements, and more visual drama could have made a huge difference.
Overall, it came very close, but the first impression wasn’t quite as strong as Claude Code.
Claude Code
Felt the most complete from the first screen
Claude Code won me over almost immediately with its hero section. That was the first thing I noticed, and it instantly felt more polished than the other two.
The layout looked sleek, the sneaker visual had the right amount of presence, and the overall composition felt like something built for an actual product launch page rather than a quick AI-generated demo.
The typography was another big reason why Claude Code stood out. The headers, section titles, and body copy all had a clear sense of hierarchy. Nothing felt randomly oversized or too similar across the page.
The design had breathing room, and the copy felt properly placed rather than dumped into sections just to check boxes. I also liked how it handled the colorways menu. It made the sneaker variants feel like a real part of the shopping experience, not just a decorative feature.
The gradient theme was present throughout the site, but it wasn’t overdone. My only gripe was the Lookbook section. It looked good, but it could have used more depth, richer details, or a stronger campaign-style presentation. Overall, Claude Code delivered the most balanced result.
Same prompt, wildly different results
This test made one thing clear: building a complete website is no longer the real challenge for AI coding tools. All three managed to follow the prompt, add the required sections, and produce something usable.
But usable is not the same as polished. Codex was fast, but it lacked refinement. Google Antigravity came impressively close with strong visuals, smart interactions, and excellent copy.
Claude Code, however, delivered the most complete package. From the hero section to typography, colorways, spacing, and overall design balance, it felt the most intentional. It didn’t just build a website — it understood the assignment like a seasoned developer.
