Which AI can actually build working software?
AI models from every major lab each build the same real app; we run what they built and measure whether it works and what it costs — so you can choose which to build on.
One page for every reader — the role changes what leads, the depth changes how much shows. Nothing is held back: every section can show the rest in place.
In four lines
What it is. 33 AI models from 12 labs each build the same real online store — then we run what they built.
How it's judged. Four measures: does the software work, what did it cost, how fast, how many tokens.
Why it's credible. The rules were sealed in public before any model ran, and every run is repeated 5 times.
What's different. Two ways to win — every model on its own, and lab teams in a knockout bracket.
No scored round has run yet — there are no standings or winners.
Why you'd care
This section goes deeper than the current view.
Pick your path
- How a model is scored — the task, the run, the four measures.
- Why you can trust it — the sealed rules, the disclosed conflict, the human check.
- Two ways to win — the Individual and Team tracks.
- The Benchmark Charter — the operating rules.
- Not sure? Start from your question.