Moonshot AI Launches Kimi K3.1 Pro, Deliberately Cut Cyber Capability

Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, supporting a 1 million-token context window. The model is live on Kimi's API at $2.00 per million input tokens and $10 per million output tokens, matching GPT-6.1 Sol and half the price of Claude Opus 5.5, with open weights scheduled for November 4, 2026. A faster variant, K3.1 Pro Highspeed, offers 2-3x faster inference at twice the price. Before release, Moonshot removed cyber-related data from mid-training and reinforcement learning after observing exploit-solving ability climb as post-training scaled. The unfiltered model solved 90.5% of tasks on Moonshot's internal xdev-bench-hard; the released version solves 81.0%, still far above the predecessor K3's 33.3%. Moonshot attributes the residual capability to general coding gains rather than targeted security training. On benchmarks, Artificial Analysis scored the model at 55, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53), trailing only Claude Opus 5.5 at 58. On Terminal-Bench 4.0 it scores 65.7%, 0.7 points behind Claude Opus 5.5 and 7.5 points ahead of GPT-6 Astra. Performance is uneven: it trails on Terminal-Bench Science (38.1 vs GPT-6 Astra's 68.1) but leads on OfficeQA Pro v2 (64.4 vs Claude Opus 5.5's 62.2). Finance-related hallucinations are down 65% compared to prior versions, and internal tool-use hallucination dropped from 23.4% on K3 to 8.3%, though GPT-6 Astra leads that measure at 5.5%.