The new model cuts cache read prices by 75%, with savings reaching up to 45% for complex agentic coding tasks