The difference in results between the high-skill and low-skill CMIP6 models highlights an urgent need to examine why some models work well and some don’t, and, ultimately, improve those with weaknesses.