I mean they could just be routing known benchmark questions (which all of SWEBench are) to a full-performance variant.