AI Game benchmarks
23 game tasks where every frontier AI model gets the same one-shot prompt. Live, playable demos. Real 0–10 scores from Julian Goldie.
What I'm testing in the Game category
Game tasks are the most useful AI model test on the bench, in my opinion. Building a playable game in one shot exposes whether the model can do three things at once: handle geometry/physics/input properly, structure a coherent game loop, and not silently fail in a way that ships a 'works but isn't a game' page. The gap between gold and bronze on game tasks is enormous — the gold demos genuinely play; the bronze demos render but you can't do anything in them.
Every Game task on the bench
23 tasks, 676 total demos across all models. Click any task to see how every AI model handled the same prompt — side by side, live and playable.
How I score Game tasks
Same three axes as the rest of the bench: runs (does the .html open to a working page), hits the brief (is the thing I asked for what came back), looks good (visual polish, motion, attention to detail). 0–10 each, averaged. Highest score on each task earns gold; second silver; third bronze. Models without a 0–10 verdict are listed as unranked on the leaderboard.
Source guides for the Game category: see the methodology page for full data provenance.
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.