Chinese lab Moonshot AI has released Kimi K3.1 Pro, a refined version of Kimi K3 that went through extra training on more compute. It scored 55 points on the Artificial Analysis Intelligence Index, 11 more than its predecessor, landing in fourth place and ahead of every OpenAI model.

Only Claude Opus 5.5 with 58 points and Claude Sonnet 5.5 and Muse Wonder Full with 56 points rank higher. The weights are due to be opened on November 4, which would make it the strongest open system available.

Agents and code

The biggest gains came in agentic work. On GDPval-AA the model scored 1846 Elo, 312 more than K3, and on Terminal-Bench 4.0 it reached 65.7% versus 13% for the previous version. That is close to Claude Opus 5.5 and clearly above GPT-6 Astra.

The model can work directly with Unity, Unreal Engine and Godot, applying fixes and building projects. It writes and runs scripts for Blender and CAD systems without MCP connectors, and Kimi.com gained a 3D Make feature for viewing and exporting 3D models.

Size and price

The architecture is unchanged: 2.8 trillion parameters with 104 billion active at a time, a 1 million token context window, and text plus image input. A Kimi K3.1 Pro Highspeed variant shipped alongside it and responds two to three times faster.

Prices dropped by a third, to $2 per million input tokens and $10 per million output tokens. Both versions are already available in Kimi.ai, Kimi Work, Kimi Code and via the API. The hallucination rate on an internal stress test fell from 23.4% to 8.3%, although GPT-6 Astra still does better there.

Cybersecurity

As post-training scaled up, the model got better at finding vulnerabilities. Before the open-weight release, security-related material was removed from the training data, and the xdev-bench-hard score dropped from 90.5% to 81.0%. Moonshot argues that open models should not push the frontier of offensive capabilities.

Commentary

Judging by this ranking, the gap between open and closed models is shrinking to a few points. For people building websites and bots, it means a strong model for agents and code may soon run on their own infrastructure, with no dependence on someone else's API. It is worth waiting for the open weights and testing the model on real tasks rather than relying only on the vendor's charts.

Source: 3DNews.