Swarm arena · round 1
Three models built the same game. Which is which?
Opus 5.5, Sonnet 5.5 and GPT-6.1-Sol each got the same prompt, one run each, no re-rolls and no edits. The games are shuffled into A, B and C. Play all three, pick the one you like best, and then you'll see who made what.
The exact prompt all three got
Make a small arcade game called "Swarm" in a single self-contained file, index.html. No libraries, no external images, fonts or sounds, and no network requests. Everything is drawn in code. The player dodges a swarm of falling locusts. It should get harder over time, keep score, show a game-over screen and let you restart instantly. It must be playable on a desktop with the keyboard AND on a phone with touch. It must fit any screen size without scrolling. Make it fun to play for a minute and good to look at. You decide everything else.
Pick your favorite
One vote each. The reveal comes right after.
Play all three to vote · 0 of 3 played
The reveal
Run your own blind compare, with every teammate on the model you pick. Try Locust here →How it was run
- One prompt, sent once to each model, word for word. Each started in a fresh empty folder with no other context.
- All three ran on High effort, in the same mode, on the same build of Locust (0.570.0).
- No re-rolls, no follow-up prompts, no fixes. If a game has a flaw, the model shipped it.
- One model hit its usage limit 11 seconds into the first try, before it wrote anything, so it ran alone about three hours later, on the same prompt, settings and build.
- The files you play are byte-for-byte what each model wrote. SHA-256:
A 4cd1d8473633ab979c0ccc543a123d5be43ee4efce8882eb4d587bd00996c4c1
B 5ef584d845a3b8aee50a4b8b796ebb8286bfac525a3908f9d3deb761c414a584
C 951f9306a202cfc6785b7431a181bf5ec8d11bcdfc718a61786ec14c0443f2af - The model that ran alone was GPT-6.1-Sol (Game B). Build times are as Locust showed them.