Curious how cheaper models would fare. Have you thought about testing gpt5-mini and similar? And even the very small qwen-coder-30b.