A Grand Tour of AI Models

In the fast-evolving world of AI-powered coding, seeing is believing. I recently took a deep dive into VizHub’s “Edit with AI” feature, a tool that lets you iterate on code with simple text prompts. The goal? To see how a diverse lineup of prominent AI models handles a range of creative and technical coding challenges. From generating vibrant data visualizations to crafting a retro pong game from scratch, this grand tour of AI capabilities revealed some surprising strengths and a few notable quirks.

The Newcomer: Kimi K2

First up was the brand-new Kimi K2. With a prompt to make a visualization “more colorful,” I was curious to see if this new model could even handle the edit format. To my surprise, it not only worked but dove headfirst into creativity, hardcoding a rainbow of colors and even introducing eye-catching gradients and a glow filter. What’s more, it generated documentation alongside the code, a fantastic bonus for understanding the changes. The result was a “pretty wild” and visually impressive transformation.

The Reliable Workhorse: Claude Sonnet 4

Next, I turned to the go-to model, Claude Sonnet 4, known for its consistent and reliable performance. The challenge? To create a fractal tree, a task significantly different from the original visualization. Sonnet 4 handled the request with aplomb, generating the necessary tree data, and impressively, it even added a smooth transition and cleaned up the old data file. The outcome was nothing short of “amazing.”

The Creative Mind: Claude 3.7 Sonnet

For a test of pure creativity and aesthetic sense, I tasked Claude 3.7 Sonnet with a fun project: “make a vanilla HTML retro pong game.” This model is reputed for its creative flair and understanding of retro styling, and it did not disappoint. Not only did it generate a clean, functional pong game, but it also went the extra mile by adding sound effects. The result was a fully playable and enjoyable retro game, showcasing the model’s impressive creative and technical abilities.

The Economical Option: DeepSeek Chat V3

In the world of AI, cost is always a factor. DeepSeek Chat is a super cheap model to run, but as the saying goes, you often get what you pay for. I gave it a simple task: “make the circles into rectangles.” While it’s incredibly fast, its output can sometimes be hit-or-miss. In this instance, it successfully made the change, proving its utility for straightforward, simple edits where speed and cost are the primary concerns.

The Powerhouse with a Price: Gemini 2.5 Pro

Gemini 2.5 Pro is known for its powerful reasoning capabilities, but this power comes at the cost of time and, well, cost. My initial simple prompt to “use purples” resulted in a long wait and, disappointingly, a failed attempt that still cost 11 cents. A more specific follow-up, “use only purples no blues or browns,” yielded a much better result, with the model even listing the color names it was using. This experience suggests that for a model this powerful, clear and precise prompts are key to unlocking its potential. In contrast, the faster and more affordable Gemini 2.5 Flash handled the “use purples” prompt more effectively on the first try.

The Grok Duo: Masters of Vanilla HTML

The Grok models from xAI, Grok 3 and Grok 4, have a reputation for excelling at generating clean, vanilla HTML. I put this to the test by asking Grok 3 to create a “3D starfield.” The result was a sweet and functional 3D animation.

Taking it a step further, I challenged Grok 4 to make the starfield animate as if you were flying through it, all while using vanilla HTML and removing the package.json file. Despite a long wait time that tested my faith, Grok 4 delivered a “brilliant” and “beautiful” 3D animated starfield with remarkably clean and concise code—under 100 lines for such an impressive effect.

The Industry Standards: OpenAI’s GPT Models

No AI model tour would be complete without OpenAI’s offerings. To test their utility beyond pure code generation, I prompted GPT-4 to “update the readme to explain the math of all this.” The model produced a wonderfully detailed explanation, turning the README into an on-demand textbook that broke down the concepts of perspective projection and animation frames. This highlighted the feature’s potential as a learning companion.

Finally, a simple test for the model that powers ChatGPT, GPT-4o: “add more stars.” The edit was fast, and the result was a denser starfield, confirming that this popular model integrates seamlessly with the “Edit with AI” feature.

A Spectrum of Capabilities

This journey through VizHub’s “Edit with AI” feature and its diverse model support showcases a vibrant and varied landscape of AI-powered development. Here’s a quick rundown of the contenders:

  • Claude Sonnet 4: The reliable default, great for consistent, high-quality results.
  • Kimi K2: A promising and affordable newcomer with a creative spark.
  • Claude 3.7 Sonnet: The go-to for creative and aesthetically pleasing projects.
  • DeepSeek Chat V3: A budget-friendly option for quick and simple edits.
  • Gemini 2.5 Pro: A powerful reasoning engine that requires patience and precise prompting.
  • Gemini 2.5 Flash: A speedy and effective alternative for less complex tasks.
  • Grok 3 & 4: Impressive performers, particularly for clean, vanilla HTML.
  • OpenAI GPT-4 & GPT-4o: Excellent for both code generation and documentation, making them valuable learning tools.

Conclusion

The power to choose the right AI for the right task is a game-changer. I encourage you to try out these models in VizHub and see which ones best fit your workflow and creative vision. The future of coding is collaborative, and with these AI partners, the possibilities are endless. Let me know which models work best for you!