As far as we know, it was "just" scale on depth (model capability) and breadth (multiple agents working at the same time).