Both finished. I didn't like either game.
I wanted to see what Claude and Codex would do with a game-building task. I picked Subway Surfers as the reference because I knew the game and wanted to try something familiar.
By the end, I had two builds to open. In the recording, I read out 19 minutes for Claude and 16 for Codex. That sounds like an easy comparison until you watch me try to play them.
With one, I was asking how to lose so I could make it stop. With the other, I was struggling to do anything. My verdict at the time was pretty simple: I didn't think either result was good.
The interesting part was seeing the difference between getting a finished response and getting a game I actually wanted to play.
What I gave each one
I used the goal workflow available in my setup at the time. The idea was to let each coding assistant keep working toward a defined result rather than stopping after its first attempt.
I had prompts prepared for both. Claude had written its own prompt and Codex had written its own prompt. I showed them side by side before starting.
That matters if you're watching this as a comparison. It wasn't an identical-prompt benchmark. The image tools also differed: Codex had image generation built into its environment, while I had given Claude access to a Google image-generation API. Later in the video, I said Claude hadn't been able to use that access.
You can see the setup from 1:30. These are the tools and results from that recording, not a claim about which current model wins every game-building task.
The part I still had to do myself
I still had to open the builds and see whether they worked as games.
I liked that Codex had generated pictures. That didn't settle whether the controls worked well or whether playing the result was any fun. A completion message couldn't answer those questions for me.
If I run this again, I'd make the acceptance test much more concrete: start the game, move the character, hit an obstacle, lose, and restart. Then I'd ask both assistants to show those steps working with the same brief.
That's a change I'd make for another test, not something I proved in this one.
The full comparison includes the prompts, the finished builds and my first reactions to playing them.
Tools
- Claude: one of the coding assistants in the comparison.
- Codex: the other coding assistant.
- Subway Surfers was the game reference. This was my own experiment, not an official version or a collaboration with its creators.