The problem is they aren’t just asking the model to do the problem they force it to do it the way a precocious ape child would then make it explain its steps like the apelets. 2000 dollar compute cluster and internal models my ass. I literally get o4 to write category theory and topos proofs that compile in Agda with $5 in api calls max. You really expect me to believe it needs frontier internal models and GPU farms for advanced hs math but higher geometry that compiles is nbd for retail?