The same prompt in the same project gives different results/slightly worse results compared to Claude Code, both using Opus model.