Time to move to locally hosted open source models - only if the thinking models could perform well locally