These changes fixed some of the token issues, but the token bloat is an intrinsic problem to the model, and Anthropic's solution of defaulting to xhigh reasoning for Opus 4.7 just means you'll go through tokens faster anyways.