Independent build · AI evaluation
Can an AI build a good game in one pass?
Game Bench gives coding models the same game brief, a clean workspace, and no feedback. Every result stays playable, so you can inspect the code and decide whether the score matches the experience.
Live benchmark · updated as new runs are scored
The problem
Most model demos hide the conditions that produced them.
A polished clip tells you little when one model had a detailed prompt, another had three rounds of feedback, and a third inherited working code. I wanted an artifact test with visible constraints.
Browser games force a model to combine interaction, state, rendering, onboarding, tuning, failure states, and polish. You can tell within a minute whether the result works.
The protocol
Same assignment. Same stopping rule.
Reproducible runs · public artifacts · cost-aware comparison
-
01
Freeze the brief
Each model receives the same mechanics, platform, art, audio, and quality constraints.
-
02
Let it build
The model starts in an empty workspace. I provide no corrections, examples, or follow-up prompts.
-
03
Score the artifact
I test gameplay, prompt adherence, engineering, and art, then publish the game, rubric, cost, and score.
What I learned
A benchmark needs an opinion about useful work.
Passing tests does not make a game fun. Visual polish cannot rescue unclear controls. A cheap run can be the better product choice even when it loses on raw score.
- The leaderboard can sort each scoring pillar, not only the final index.
- The efficiency chart separates raw quality from the cost required to produce it.
- Blind routes let players judge the game before seeing the model name.
Judge the outputs
Play first. Reveal the model later.
Nine anonymous builds are available in the blind-test flow.