Yes, I feel the next differentiation in model capabilities would be wrt how 'long of a horizon' goal they could autonomously work toward without breaking. We're in need for a new generation of evaluations and benchmarks for this -- METR doesn't cut it anymore.