Carrick AI GameBench Benchmarking AI coding through games用游戏评测 AI 编程能力
Open by default

Read the code behind every score.

GameBench is not a black-box ranking site. The game briefs, browser tests, scoring rules, clean model-made source, and published result records are public.

Game rules

What every model receives

Starter code, game briefs, public state shapes, browser checks, time limits, seeds, and point values.

Scorekeeper

How the browser awards points

Real input, controlled time, game-state checks, build gates, clean rebuild verification, and score aggregation.

Model output

What you can download

Playable builds, screenshots, licenses, and clean source—with credentials, private model traces, and provider responses removed.

What you may reuse

GameBench code is Apache-2.0. Reproduce tasks retain their upstream license, exact commit, copyright notices, and distributed-material hashes. A published file never broadens the license of upstream work.

Site typefaces

The interface self-hosts Oxanium, Silkscreen, CAGB Rounded SC (derived from WenYuan Rounded SC), and Fusion Pixel under the SIL Open Font License 1.1. The exact Oxanium, Silkscreen, WenYuan Rounded SC derivative, and Fusion Pixel license texts ship with the site.

Why games open on another domain

Model-generated JavaScript is untrusted. Every game runs on an isolated play domain with outside network access disabled and no access to the main site's cookies, storage, or page.