It would be nice if it could estimate GPT usage costs with a dry-run.