Moonshot AI's Kimi K3.1 Pro Closes Gap to 3 Points Behind Claude Opus 5.5

A new reasoning model from Chinese company Moonshot AI has landed in fourth place on the Artificial Analysis Intelligence Index, scoring 55 points. That result places it just three points behind Claude Opus 5.5, which holds the top score of 58. The model, Kimi K3.1 Pro, is a post-trained version of its predecessor, Kimi K3, developed with significantly more compute for that stage of training. The upgrade resulted in an 11-point score increase, moving the model from its previous ranking to a position where it trails only Claude Opus 5.5, Claude Sonnet 5.5 (56), and Muse Wonder Full (56).

The company intends to release the model's weights by November 4, 2026. Should that schedule hold, K3.1 Pro would become the highest-scoring open-weights model available. This would narrow the gap between the leading open-weights models and proprietary models from 12 points to just 3 points, assuming no other models change their scores by then. Moonshot is also shipping a variant, K3.1 Pro Highspeed, which runs two to three times faster. Both are accessible now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

The model's biggest performance improvements are in agentic tasks. On the GDPval-AA benchmark, K3.1 Pro reaches an Elo of 1,846, a gain of 312 points over K3, placing it just behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra (58.2 percent) and close behind Claude Opus 5.5 (66.4 percent), while the previous K3 scored 13 percent. Hallucination rates on Moonshot's internal Kimi Hallu Bench Hard, a stress test rather than a measure of general error rates, declined from 23.4 percent with K3 to 8.3 percent. GPT-6 Astra performs better on this benchmark with a score of 5.5 percent.

K3.1 Pro is designed to work directly with game engines such as Unity, Unreal Engine, and Godot for tasks ranging from minor fixes to full project builds. On the BenchCAD benchmark, it scores 92.7 percent with tools, close to Claude Opus 5.5 (96.2 percent) and GPT-6 Astra (95.9 percent). Without tools, it achieves 72.8 percent, nearly matching Claude Opus 5.5's 73.0 percent. The model can write and execute scripts for Blender and CAD software without MCP connectors, and Moonshot has launched a feature on Kimi.com for previewing, editing, and exporting 3D models.

The model's specifications are unchanged from its predecessor, with 2.8 trillion total parameters, 104 billion active parameters, and a one million token context window. Token prices have dropped by 33 percent, with input costs at $2 per million tokens, output at $10, and cache hits at $0.20. Each Intelligence Index task costs $1.98, nearly identical to K3's $2.00, since the model generates 14 percent more output tokens.

For the planned open-weights release, Moonshot removed cyber-related data from both training phases because vulnerability research capabilities grew as post-training scaled up. On an internal test that reproduces recently disclosed vulnerabilities in software like the Linux kernel, Node.js, and JavaScriptCore, the version without data cleaning solves 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent, while DeepSeek-V4.1-Flash reaches 85.7 percent on the same test and Kimi K3 scored 33.3 percent. Moonshot states that no open-weights model should advance the absolute frontier of offensive cyber capabilities, though the company also wants to gradually expand what open models can do because they support defensive functions such as auditing legacy systems.