I could be wrong, but what I'm noticing about these headline-grabbing (deservedly) results from frontier models on long-held theories of computation limits is, that they are more showing that the long-held theories are approximations, not so much that they are wrong. The "Exponential Time Hypothesis" and the square-root bound for ECDLP are in this category: both the results that show they are wrong, show they are wrong to an absurdly infinitesimal extent. Maybe it's more like saying that things are fuzzier than you think, not clean, not so much saying that the general rule you have in mind is wrong. Just a guess, though.