When Model Drift Breaks Your Vibe-Coding Flow

I was “vibe-coding” with OpenClaw using Codex 5.3 to refine a little web app I’m building as a learning project. I went to bed and OpenClaw reverted to Claude Sonnet overnight. I had Sonnet set as the default to try and save on usage tokens.

I did not realize this when I started again the next day. I asked OpenClaw to make some style changes. It went off the rails and overwrote a lot of the existing styling, and reverted some changes, making a mess of the app.

I see a huge difference in the level of understanding with OpenAI GPT 5.3 Codex vs Anthropic Claude Sonnet 4.5. I will need to do more testing with Claude 4.6 which is the latest from Anthropic. I am sure it will be better.

This wasn’t a big deal with my little test app. I was able to fix it in ~15 minutes, after switching back to Codex 5.3. I have seen other horror stories of agents deleting files and the like.

It is important to test different models to find what works well. Also, it is critical to have a good backup and recovery plan for when AI does the unexpected!

Before and After