If you compare Li's IKP scores and SimpleBench scores across GPT-5.1 and before vs. 5.2-5.4, you will see random variation and often even regression with the model counter increment. Exactly the same thing happens with Opus 4-series models as well.

The reason is common: post-training quite commonly displaces obscure facts (for IKP) and spatial reasoning skills (for SB) occupying valuable "real estate" of weights/experts in favor of coding skills in commercial demand. The only alternative explanation I could think of is quantization while keeping param count the same, maybe one can devise more but the size increase you hypothesize for 5.1>5.2 seems to clearly contradict this phenomenon.

As of 30:1 sparsity for 4o, I have never suggested it. It seems reasonable to assume 4o, o1, o3, 4.1 and 5-5.4 all have around 100A2000B params, that is ~20:1 sparsity. Regarding the margins and the price of H100 hours, which time and which provider do you refer to? It varied a lot during recent years and even months. Also, I advise to treat prices very carefully and not as a source of the ultimate truth, as o1 is 7.5x costlier than o3 (or 4.1) but is clearly not very much different in size, if at all.