I ran four AI coding agents through a live snake-game battle, then did something a normal bake-off never does: I let them read each other’s code and try again. That second round is the whole point of the video, and it is where the result actually got interesting.
Round one sets the field
I gave Claude Code, Codex, Cursor, and Google Antigravity the same short prompt, “build the snake game,” live, no human edits, same timer. Models were pinned: Claude Code on Opus 4.6, Codex on GPT 5.4 High, Antigravity on Gemini 3.1 Pro, Cursor on Auto mode. None of them knew they were competing.
The first pass sorted them fast. Claude Code shipped a single polished HTML file that tracked a best score and animated when the snake ate the red block. Codex looked a touch cleaner, added a cherry, and updated the score in real time. Antigravity worked but opened on a “game over” screen with a plain background. Cursor broke the moment I hit play again. That is why I do these live. You watch the failure happen instead of reading a summary that hides it.
Round two: the battle royale
Then I raised the stakes. I handed each agent the paths to all four HTML files with one instruction: here is the code from three other agents, improve your own to beat them or keep what you have, and if your updated version does not run, you are eliminated.
Now the instincts showed. Claude Code took the risk and added a gradient background, swapped the red dot for a cherry, and kept a plus-one animation when eating it. Codex went further on motion, giving its cherry a small breathing, vibrating effect it did not have the first time, a clear sign it had read Claude’s file. Antigravity played it safe and shipped no animation at all, even with the other code in front of it. Cursor, broken in round one, recovered into a working game that looked almost exactly like Claude’s first iteration. If you cannot beat them, be them.
Playing all four back to back, my pick was Claude Code on vibes: the vibrating cherry plus the plus-one animation just felt best. Codex was a very close second, Cursor third after its fix, and Antigravity last for staying flat.
My take
I am not saying Claude is best at everything. The point is that one vague prompt produced four very different instincts, and you only see that by putting the agents side by side and letting them react to each other. The full head-to-head, including every game running, is in the two videos above.
Play all four games at cloudyeti.io/snake-battle and rank them yourself, and subscribe for more AI engineering experiments.
I specialize in helping teams pick and wire up the right AI coding agents for real work. Book a call at cloudyeti.io/meet.