A side-by-side look at which AI models can actually design.
Every few days someone drops a new leaderboard proving that Model A is 2.3% better than Model B at reasoning, coding, or being polite. Which is fine, but none of those spreadsheets tell you what you actually want to know: if I ask this thing to build a landing page, will it look good?
We asked a growing list of AI models to design five landing-page concepts for a second-brain note-taking app. Same prompt. Same brief. Multiple attempts preserved. Then we laid out the results so you can see the differences with your own eyes, no benchmark literacy required.
A gallery, not a spreadsheet. The home page groups model runs by condition: baseline, design-skill-enabled, taste experiments, and everything in between. Each card shows five real rendered previews. These are actual generated pages, not slides from a press kit.
Compare anything. Pick any two runs and look at them together in a single shareable URL. Baseline vs. design skill. Model X vs. Model Y. Iteration 1 vs. iteration 5. It is a fast way to turn "vibes say this one is better" into "okay, yeah, that one is better."
Rankings with actual notes. The rankings page is subjective, because design is subjective. We call out what worked, what felt like a template, and where a model showed real taste instead of just following instructions.
Lab Guess. Look at a generation, guess the model, see if you are right. It is surprisingly educational. You start noticing tics: the models that turn every page into a card grid, the ones that love oversized typography, the ones that quietly know how to use whitespace.
Model selection is a design decision. Some models give you solid structure and weak taste. Some produce one gorgeous screen and ignore the rest of the brief. Some completely change character when you toggle a design skill.
WhichAI.dev makes those differences obvious. It is for builders choosing a tool, researchers who want a corpus they can read without a CLI, and anyone who heard AI can build apps and wants a visual answer to "okay, but is it any good at design?"
Running fresh model generations is the main cost behind WhichAI.dev. If the gallery helps you choose a model—or you just want to see more models added—you can buy Dara a coffee to help fund the next batch.
Greptile's Open Source Program provides AI code review for the project.
OpenAI - Codex for Open Source provides tooling and credits that support open-source maintenance and benchmark development.