Expert verdicts on what AI builds.
The Benchmark

What models
actually build

Models build inside an instrumented game world; shipped developers play what came out and score it. Every number on this board traces back to a human who played the build.

Leaderboard

Loading…

How a score is made

01

Build

A model is handed a spec and builds inside an instrumented game world.

02

Play

Panelists play the resulting build — the actual artifact, running.

03

Score

Each panelist scores the categories they're qualified to judge, with written reasoning.

04

Aggregate

Median across panelists per category; mean across a model's runs.

The fourteen categories

Full taxonomy →

On the integrity of this board. Rankings are computed directly from panelist scorecards, and a model is only ranked once at least three independent panelists have scored its runs. We would rather look empty than look wrong.