ChinaMoonshot AI’s 3-trillion-parameter model costs 20% of GPT-6 Astra.

It ranks fourth on Artificial Analysis, first in China and ahead of GPT-6 Astra.

Moonshot says large-scale reinforcement learning across hundreds of downstream domains lifted its Artificial Analysis score 11 points over Kimi K3.

Its 65.7% on Terminal Bench 4.0 still trails Claude Opus 5.5’s 66.4%.

Moonshot will publish the open weights by 4 November.

How each outlet framed it
Zhidx
tests Kimi K3.1 Pro hands-on: a CSS-3D Rubik's cube built in about 10 minutes, a Godot game in about 20, and prices 20% of GPT-6 Astra's

Sources: Zhidx