“Removing repeated instructions from prompts for GPT-5.6 raised scores by 10-15% and cut tokens by up to 66% in OpenAI's tests.”