Terminal-Bench result: DeepSeek V4-Pro outperforms GPT-5.5 in long-horizon agent tasks