It’s cheap to copy already built models from their outputs, but likely still expensive to train new models that push the boundaries.