“On the DeepSWE coding benchmark, KIMI K3 scored 67.5, which is 8.5 points ahead of Opus-4.8 and 0.5 points ahead of GPT-5.5.”