Kimi K3.1 Pro is out. Not a new model. It's K3 run through RL again with heavier post-training compute, and it's within a few points of the top closed models on coding, 3D, and finance. Terminal-Bench 4.0: 65.7%, over GPT-6 Astra's 58.2%, 0.7 points short of Claude Opus 5.5. On Kimi's own long-horizon eval, K3 went from 55.95% to 61.2%. Third-party testers also flagged the bill. It eats tokens, and a single task takes about 2x what Opus High needs.
Kimi K3.1 Pro is out. Kimi K3.1 Pro is out. Kimi K3.1 Pro is out.