The "efficiency" meantioned in blog post you have linked is the price difference between Deepseek and o1, it doesn't mean that GPT-5 or other SOTA models are less efficient.