Kimi K3.1 Pro scores 55, three points behind top AI models
Moonshot AI's new reasoning model cuts API costs by a third while outperforming its predecessor on agentic and financial tasks.
By Jisoo Vega ·
Consumers accessing artificial intelligence services through Moonshot AI's platforms can now use Kimi K3.1 Pro, a model that scored 55 points on the Artificial Analysis Intelligence Index. This result places the system in fourth position globally, three points behind the leading proprietary systems. The price for API access has dropped by 33 percent, with input costs set at $2 per million tokens and output costs at $10 per million tokens.
The new model represents an 11-point improvement over its predecessor, Kimi K3, which previously scored 44. According to Artificial Analysis, only Anthropic's Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full currently hold higher scores. K3.1 Pro is a post-trained version of K3 that utilized substantially more compute resources during its development phase. Moonshot plans to publish the model weights by November 4, 2026. If released as scheduled, it would become the highest-scoring open-weights model available, narrowing the performance gap between open and proprietary systems from 12 points to 3.
Performance gains were most evident in agentic tasks. On the GDPval-AA benchmark, K3.1 Pro reached an Elo rating of 1,846. This marks a 312-point increase from K3 and places the model just behind Claude Opus 5.5, which scored 1,866, and Muse Wonder Full, which scored 1,857. On Terminal-Bench 4.0, the new model achieved a score of 65.7 percent. This surpasses GPT-6 Astra, which recorded 58.2 percent, and falls slightly short of Claude Opus 5.5, which scored 66.4 percent. By comparison, K3 managed only 13 percent on that specific test.
Moonshot reports a sharp reduction in hallucinations for the new system. On its internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent with K3 to 8.3 percent with K3.1 Pro. The company describes this benchmark as a stress test rather than a general error rate measure. It notes that GPT-6 Astra still leads this specific metric with a score of 5.5 percent. In internal finance benchmarks, the model gained 7.2 points on Financial Modeling and 11.6 points on Financial Research. Moonshot states that K3.1 Pro tops all tested models on OfficeQA Pro v2.
The model supports 3D modeling and coding tasks without requiring MCP connectors. It works with game engines including Unity, Unreal Engine, and Godot, and can write and run scripts for Blender and CAD software. On the BenchCAD benchmark, K3.1 Pro scored 92.7 percent with tools. This compares to 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra. Without tools, the model reached 72.8 percent, slightly below Claude Opus 5.5's 73.0 percent. Moonshot has added a 3D Make feature on Kimi.com to allow users to preview, edit, and export 3D models.
The architecture remains unchanged from K3, featuring 2.8 trillion total parameters with 104 billion active parameters. The model supports a one-million-token context window and handles both text and images. Artificial Analysis places K3.1 Pro on the top of the Pareto frontier, indicating that no model of comparable size scores higher on its index. Because the model generates 14 percent more output tokens, the cost to run each Intelligence Index task is $1.98, nearly identical to K3's $2.00 cost.
Moonshot included a section on cybersecurity in its technical blog post. The company states that vulnerability research capabilities improved as post-training scaled up, since cyber-related content was part of the general mid-training corpus. For the open-weights release, Moonshot removed cyber-related data from both training phases. On its internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software like the Linux kernel and Node.js, the uncleaned model solved 90.5 percent of tasks. After cleaning, the score dropped to 81.0 percent. DeepSeek-V4.1-Flash scored 85.7 percent, while Kimi K3 scored 33.3 percent.
Moonshot says it does not believe any open-weights model should push the absolute frontier of offensive cyber capabilities. The company aims to raise the capabilities of open models gradually, noting that they also support defensive work such as auditing legacy systems. Anthropic previously demonstrated that GLM-5.3, which scores 45 on the Artificial Analysis index, can write exploits at levels close to Claude Mythos.
Source: The Decoder
