I Made Two AIs Compete to Build a Full Among Us in Godot — Then I Swapped Models Mid-Run

The Setup

I wanted to answer a question honestly: are open-weight models good enough for a real project yet? Not a benchmark — an actual, complete thing you can play.

So I ran a duel: same prompt, zero human intervention, pure agentic coding. The goal: a full Among Us clone in Godot 4.4 — 8 characters, 6 rooms, real tasks, emergency meetings with voting, one AI-controlled impostor.

The contenders: Claude Code vs an open-weight model running on OpenCode.

The Twist

The open-weight model got stuck in a repetitive error loop about halfway through. Since OpenCode is model-agnostic, I did something that's simply impossible with a closed API: I swapped the model mid-run and kept going — same tooling, same workflow, same repo.

That moment is the real thesis of this post: open weights don't just mean free inference, they mean exit options. If your closed-API model stalls, you don't swap models — you restart your project.

Honest Results (fun part first)

🎯 Claude: came out of the gate with tests, versioning and a full game in one pass. It even had some fun with it — the impostor killed me the moment I left the cafeteria. On camera. I died with everyone watching.

🧩 Open weights: creative but fragile. It got bogged down in error loops; the mid-run swap saved the experiment, but the final build ended well below the closed model.

🔥 Bonus chaos: my local machine ran out of RAM mid-recording. All left unedited — the suffering is part of the documentary.

Watch the full run

The whole 25-minute run is on YouTube, in Spanish with all the failures intact:

Try it yourself

The full prompt and resulting code are free and open source:

👉 github.com/bramenn/among-us-ai

Fork it, run your own duel, and tell me what happened — which model stalls first?

Discussion

What's your daily stack: closed API, open weights, or a mix? And has anyone here actually swapped models mid-project because the first one stalled — or are we all just silently re-rolling prompts?

If you want me to run a rematch with a specific open-weight model, drop the name in the comments 👇