Can AI write code? In October 2026 the useful version of that question is narrower: can a model write the specific thing you need, from one prompt, without someone fixing it afterwards? This shelf answers it with finished builds rather than opinions. Every page here asks one concrete question, such as whether AI can build a website, an app or a playable game, and then shows what four frontier models actually produced when they tried.
The short answer from our first four briefs is yes, with caveats you can see, and the four newer briefs, a phone app, a MOBA, an invented brand and a kanban board, point the same way:
- All four models finished the landing page, the dashboard and the physics simulation inside the time limit, with zero runtime errors recorded on those builds.
- The 3D shooter was harder. Every model produced a playable game, but Claude Opus 5.5 was still working when its 150-minute limit arrived, and two of the four recordings end in a game over before the scripted input runs out.
- Nobody needed a framework. Every build is a single HTML file, with a few extra script files for the games, that opens straight from disk.
The methodology explains what one attempt means here and how interruptions are handled. If you want to try a brief on your own favourite model, the exact prompt is on every brief page, with a copy button.