Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0. GPT-6 Astra manages 58.2%. Claude Opus 5.5 is still ahead at 66.4%, by about a point. It carries a 1M context window and the full weights go up on November 4. They also stripped cyber-attack training data out of both the mid-training corpus and the RL mix before release, which means the copy you download is deliberately worse at finding holes in your own servers. Fair trade. Read the number twice anyway.