China's Kimi K3.1 Pro comes within three points of Claude Opus 5.5, with open weights promised by 4 November

Chinese company Moonshot AI has released Kimi K3.1 Pro, a reasoning model built on Kimi K3 with much more post-training. Independent benchmarker Artificial Analysis places it fourth on its Intelligence Index with 55 points — 11 more than K3 and three behind the leader, Claude Opus 5.5 (58), according to The Decoder.

Key facts

• Size: 2.8 trillion total parameters, 104 billion active; a 1-million-token context window; text and image input.
• Price: about 33% cheaper than K3 — $2 per million input tokens and $10 per million output tokens. A faster “Highspeed” variant costs twice as much.
• Open weights: Moonshot says it will publish the full weights by 4 November 2026. If it does, K3.1 Pro would become the highest-scoring open-weights model on the index.
• Gains: the biggest improvements are in agentic tasks, coding and 3D/CAD modelling; the model can work with Unity, Unreal Engine, Godot and Blender, the company says.

Benchmarks (company-reported unless noted)

On Terminal-Bench 4.0 Moonshot reports 65.7%, close to Claude Opus 5.5 (66.4%). On Artificial Analysis’s GDPval-AA, it reaches 1,846 Elo versus 1,866 for Opus 5.5. Moonshot says hallucinations on its internal stress test fell from 23.4% to 8.3%. Vendor benchmarks should be treated with caution until independently reproduced.

Cyber limits

Moonshot says it removed cyber-related training data for the open release, because vulnerability-research ability kept growing with training. On its internal exploit benchmark the cleaned model solves 81.0% of tasks versus 90.5% for the uncleaned version — so the model remains highly capable in this area.