The lede
Kimi K3.1 Pro and K3.1 Pro Highspeed
Moonshot AI · made the Sunday
Moonshot is claiming 65.7% on Terminal-Bench 4.0 against Opus 5.5's 66.4% for a 3T-class model with a million tokens of context, and says the full weights ship by November 4 — which would put something within a point of the frontier on agentic terminal work onto hardware you can rent. Read the post for the engineering rather than the scoreboard: the claimed gains come from RL infrastructure (sandboxing, rollout scheduling, environment quality checks) applied across hundreds of domains rather than any new architecture, and the most concrete disclosure is a deliberate capability cut — stripping cyber data from mid-training and RL dropped their internal xdev-bench-hard pass@3 from 90.5% to 81.0%. All of it self-reported, and all of it checkable in eleven days.