See which AI builds the better game.
A playable benchmark for spatial, temporal, and causal coherence in AI-built worlds.
July 2026 pilot round
24 playable games from 8 models, built from the same 3 briefs.
Evidence you can inspect.
A single score rarely matches what you notice using a model—whether a world stays coherent over time, understands space, preserves state, and produces sensible consequences when you act. A playable game tests those qualities together, so the build itself becomes the evidence.
Shared briefs
Every model receives the same game task, so the playable output stays at the center of the comparison.
Playable outputs
The games are the evidence. Open any build, test the controls, and compare the same task side by side.
Blind evaluation
Arena voters play anonymous pairs before model identities are revealed. Ratings publish only after enough comparisons.
Open artifacts
Round JSON, run context, captures, and the open harness make each published claim inspectable.
Every model receives the same game task, so the playable output stays at the center of the comparison.
The games are the evidence. Open any build, test the controls, and compare the same task side by side.
Arena voters play anonymous pairs before model identities are revealed. Ratings publish only after enough comparisons.
Round JSON, run context, captures, and the open harness make each published claim inspectable.
Scoring thresholds, run disclosures, and the honesty rule live on the methodology page.
Play blind. Then decide.
Play two anonymous AI-built games, choose which one you would keep playing, then see which models made them. Enough blind votes set each round’s Arena Rating.
Frequently asked questions.
What is WorldBuild Bench?+
WorldBuild Bench is a playable benchmark for spatial, temporal, and causal coherence in AI-built worlds. A single score rarely matches what you notice using a model—whether it stays coherent over time, understands space, preserves state, and produces sensible consequences when you act. Games put those qualities in one testable world: models build playable browser games from shared task briefs, and humans compare the results in the WorldBuild Arena. It evaluates the worlds models build, not only the answers they produce.
How is WorldBuild Bench scored?+
The public quality signal is the Arena Rating: anonymous pairwise human evaluations fit with a Bradley-Terry model and displayed on an Elo-like scale. Ratings remain unpublished until a round clears its comparison threshold. Build status, generation time, cost, and code size are context rather than quality scores.
What is the WorldBuild Arena?+
A blind pairwise evaluation: play two anonymous AI-built games for at least 20 seconds each, choose which one you would rather keep playing, then see the model identities. Named model comparison lives on each round page.
Is the benchmark open source?+
Yes. The WorldBuild TypeScript harness is published on GitHub under MIT, round data is downloadable as JSON, and the methodology explains how every public metric is produced.
When does a model get a public rating?+
A model appears with an Arena Rating once its published games have enough counted human comparisons. Before that, the round page shows the generated builds and artifact data with an awaiting-evaluations state.
How often are rounds run?+
Roughly monthly, timed to significant model releases. Each round uses fresh task briefs to prevent training-data leakage. The first round was published on July 13, 2026.