Discovered this while I was researching about why some LLM models perform significantly worse at times. I have experienced this more with Claude.