Which might be a side-effect of the reasoning. In my experience whenever these models solve a math or logic puzzle with reasoning, they generate extremely long

Which might be a side-effect of the reasoning. In my experience whenever these models solve a math or logic puzzle with reasoning, they generate extremely long and convoluted chains of thought which show up in the solution. In contrast a human would come up with a solution with 2-3 steps. Perhaps something similar is going on here with the generated code.