Maybe use fixed bugs from closed source to benchmark these "frontier" models.