Moonshot AI Posts Kimi K3.1 Pro, Cuts API Rates and Trims Cyber Data for Open Release
The model's claim-level hallucination rate fell from 23.4 percent to 8.3 percent, and Moonshot plans to publish weights on Nov. 4.
By Amina Kaminski ·
Moonshot AI released Kimi K3.1 Pro on Oct. 23, 2026, a reasoning model that scores 55 on the Artificial Analysis Intelligence Index and places fourth among all models tested. The score is 11 points above the prior Kimi K3 and three points below Anthropic's Claude Opus 5.5, which scores 58. Claude Sonnet 5.5 and Muse Wonder Full each score 56.
The model is a post-trained version of K3 and carries the same architecture: 2.8 trillion total parameters, 104 billion active parameters, a one-million-token context window, and text and image support. API prices drop 33 percent, with input at $2 per million tokens, output at $10 and cache hits at $0.20. Because K3.1 Pro generates 14 percent more output tokens, the cost of a single Intelligence Index task is $1.98, against $2.00 for K3.
Moonshot plans to publish model weights by Nov. 4, 2026. Artificial Analysis said that if the release proceeds, K3.1 Pro would become the highest-scoring open-weights model, cutting the gap between open and proprietary systems from 12 points to 3. Artificial Analysis also places the model on the top of the Pareto frontier, meaning no model of comparable size scores higher on its index.
The largest gains appear in agentic task benchmarks. On GDPval-AA, K3.1 Pro reaches 1,846 Elo, which Artificial Analysis puts 312 points above K3 and just below Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Terminal-Bench 4.0, Moonshot reports 65.7 percent, against 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5; K3 scored 13 percent on the same test.
Moonshot also reports changes in its own evaluations. On the Kimi Hallu Bench Hard, claim-level hallucinations fell from 23.4 percent with K3 to 8.3 percent. The company describes that benchmark as a stress test rather than a general error-rate measure, and said GPT-6 Astra scores 5.5 percent on it. On internal finance benchmarks, the model gained 7.2 points on Financial Modeling and 11.6 points on Financial Research, and Moonshot said it leads all tested models on OfficeQA Pro v2.
For the open-weights release, Moonshot removed cyber-related data from both training phases. On xdev-bench-hard, an internal test that reproduces recently disclosed vulnerabilities in software including the Linux kernel, Node.js and JavaScriptCore, the uncleaned model solved 90.5 percent of tasks against 81.0 percent after cleaning. DeepSeek-V4.1-Flash scored 85.7 percent and Kimi K3 scored 33.3 percent. Moonshot said no open-weights model should push the absolute frontier of offensive cyber capabilities, while also noting that open models support defensive work such as auditing legacy systems.
K3.1 Pro works with game engines including Unity, Unreal Engine and Godot, and can write and run scripts for Blender and CAD software without MCP connectors. On BenchCAD it scores 92.7 percent with tools, against 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra. Moonshot has added a 3D Make feature on Kimi.com for previewing, editing and exporting 3D models. A variant called K3.1 Pro Highspeed runs two to three times faster, the company said. Both versions are available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
