"Which AI for coding?" has no single answer, because a model that writes the best landing page is not automatically the one that ships the most stable game. This shelf answers the question one job at a time: websites, web design, user interfaces, apps, games and general coding. Each page is built around the same evidence, four frontier models given one identical brief, one attempt, no edits, and every claim on it points at a build you can open.
A quick map of what the builds have shown so far:
| If your job is… | Watch this brief first | What tends to separate the models |
|---|---|---|
| A marketing site | Landing page | Brand personality versus template polish |
| An internal tool | Analytics dashboard | Chart craft, table details, dark mode |
| Physics or simulation | Spinning hexagon | Collision handling, whether the pile settles |
| A 3D game | First-person shooter | Atmosphere versus finishing inside the time limit |
| A phone app | Finance app | Native feel: transitions, feedback, safe areas |
| A strategy game | One-lane MOBA | Game systems, AI opponents, procedural art and sound |
| A brand from scratch | Open brief | Invention and taste when nothing is specified |
| A team tool | Kanban board | Drag and drop, undo, state that stays consistent |
The figures behind the pages are measurements, never scores: wall time, output tokens, file size, runtime errors and attempts, taken from the run logs. Prices come from each vendor's own published price list, with the date we checked. When a page says one model is the better pick, it says why, and the build is one click away so you can disagree.
If you only have a minute, skip the reading: the blind test hides the model names and asks you to pick first.