Better models require less prompt engineering per task, but they also unlock higher-value results that sophisticated prompting can reach ...