I Built a Web Dev Arena to Test AI Coding Models Side by Side

Every AI coding model has a benchmark score. None of them tell you what actually matters: does the output look good? Is the UI responsive? Do the interactions feel right? SWE-bench measures whether a model can patch a GitHub issue. It does not measure whether the todo app it builds has proper drag-and-drop, or whether the landing page it generates looks like a real product vs. a homework assignment.

That gap is why I built the Web Dev Arena. I wanted to see what happens when you give 6 different AI models the exact same prompt and compare the raw HTML output side by side. Not synthetic benchmarks. Not cherry-picked examples. The same 10 tasks, the same system prompt, rendered in iframes next to each other so you can interact with every implementation yourself.

How It Works

The setup is simple. Each model gets a system prompt: "You are an expert web developer. Generate a complete, self-contained HTML file with inline CSS and JavaScript." Then it gets the task description. The output is a single HTML file. No frameworks, no build step, no external dependencies (except CDN links like Three.js when the task calls for 3D). Every model gets the same prompt word for word.

The 10 tasks span a range of difficulty. Simple ones like a snake game and a todo app with drag-to-reorder. Medium tasks like a split-pane markdown editor, a weather dashboard with CSS-animated icons, and a SaaS landing page using a specific design system. Complex tasks like a 3D Golden Gate Bridge scene and an interactive solar system with all 8 planets in Three.js. The arena UI lets you pick a task, toggle which models you want to compare, and see them rendered in side-by-side iframes. You can open any implementation full screen to interact with it directly.

What Surprised Me

Composer 2 and Kimi K2.5 both completed all 10 tasks. Droid (running Claude Sonnet 4.6 under the hood) also hit 10/10. MiniMax M2.5 got 9 out of 10. But completion rate only tells half the story.

The more interesting finding was how much the outputs differ in craft. Same prompt, wildly different results. One model's calculator has perfectly aligned buttons with subtle hover states and keyboard support. Another model's calculator technically works but looks like it was styled in 2004. One model's particle wave animation runs at 60fps with smooth mouse repulsion physics. Another model's version stutters and the particles cluster in the corners.

MiniMax was the biggest surprise. It is not a model most developers have heard of, but its outputs consistently had strong visual design. The landing pages looked polished. The weather widget had thoughtful layout choices. For a model running on the Anthropic-compatible API at a fraction of the cost, the quality-to-price ratio is hard to beat.

Kimi K2.5 was another standout. It is on an unlimited plan, which means you can run it on high-volume tasks without watching a usage meter. The code quality was clean, the UIs were functional, and it handled the complex 3D tasks without choking. For a model that most people outside of China have not tried, it consistently punched above expectations.

What Separates "Working" from "Good"

After reviewing 50+ implementations across all models, patterns emerged. The best outputs share a few traits that the weaker ones lack:

Proportional spacing. Good implementations use consistent padding and margins. Bad ones dump elements on the page with random gaps. This is the single biggest tell. If the model understands visual rhythm, everything else tends to follow.

Interaction polish. Hover states, focus rings, transitions, keyboard support. The best implementations feel like someone actually used the app and thought about the experience. The worst ones render static HTML that happens to have a click handler.

Constraint adherence. The prompts specified a design system: cream background, black borders, pill-shaped buttons, pink accent color. Some models nailed this. Others ignored half the constraints and generated their own color scheme. Following instructions is itself a signal of model quality.

Progressive enhancement. The best snake game implementations have a start screen, score tracking with localStorage, game over with replay, and mobile touch controls. The weakest ones just render a grid and call it done. The prompt asked for all of these features. Only some models delivered all of them.

Try It Yourself

The full arena is live at demos.developersdigest.tech/arena. Pick a task, select your models, and compare. Every implementation is interactive. You can play the snake games, type in the markdown editors, drag todos around, orbit the 3D scenes.

If you are evaluating which AI coding model to use for frontend work, this is more useful than any leaderboard. Benchmarks measure capability in the abstract. The arena shows you what the model actually builds when you ask it to build something.