Newer Claude models use more tokens but cost less per task solved