My purely unfounded, gut reaction to Opus 4.7 being released today was "Oh, that explains the recent 4.6 performance - they were spinning up inference on 4.7."

Of course, I have no information on how they manage the deployment of their models across their infra.