Moonshot AI’s Kimi K3.1 Pro Scores 55 Points on Intelligence Index

Moonshot AI has released Kimi K3.1 Pro, a reasoning model that has climbed to the fourth position on the Artificial Analysis Intelligence Index with a score of 55. This result places the new system just three points behind the leading proprietary models, marking a significant shift in the competitive landscape for artificial intelligence capabilities. The release represents a post-trained evolution of the previous Kimi K3 architecture, designed to leverage significantly more post-training compute for improved performance across various technical domains.

The numerical improvements are substantial compared to the predecessor. Kimi K3.1 Pro achieves an 11 point increase over the original K3 model. According to Artificial Analysis, only three models currently rank higher: Claude Opus 5.5 at 58 points, Claude Sonnet 5.5 at 56 points, and Muse Wonder Full at 56 points. If Moonshot adheres to its plan of releasing the model weights by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available. Such a move would reduce the performance gap between open and proprietary models from twelve points to just three, assuming no other major releases occur in the interim.

The model is already available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Moonshot is also shipping a variant named K3.1 Pro Highspeed, which the company states runs two to three times faster than the standard version. The primary performance gains for the new release are observed in agentic tasks. On the GDPval-AA benchmark, the model reaches an Elo rating of 1,846, a 312 point increase over K3. This places it slightly behind Claude Opus 5.5, which scores 1,866, and Muse Wonder Full, which scores 1,857.

Coding capabilities have also seen marked improvement. On Terminal-Bench 4.0, Moonshot reports a success rate of 65.7 percent. This figure exceeds the 58.2 percent achieved by GPT-6 Astra and trails Claude Opus 5.5 by less than one point. In contrast, the previous K3 model managed only 13 percent on the same test. Hallucination rates have also dropped sharply. According to the company, claim-level hallucinations on the internal Kimi Hallu Bench Hard decreased from 23.4 percent with K3 to 8.3 percent with the new model. While GPT-6 Astra maintains an edge at 5.5 percent on this specific stress test, Moonshot notes that this benchmark is designed to push the limits of error rates rather than measure general accuracy.

In specialized professional tasks, the model shows distinct advantages. On internal finance benchmarks, K3.1 Pro gained 7.2 points in Financial Modeling and 11.6 points in Financial Research. For general office tasks, Moonshot claims the model outperforms all tested competitors on OfficeQA Pro v2. The system also demonstrates strong capabilities in three-dimensional modeling and coding. It can interact directly with game engines such as Unity, Unreal Engine, and Godot to handle both minor fixes and full project builds. On the BenchCAD benchmark, the model scores 92.7 percent when using tools, close to the 96.2 percent recorded by Claude Opus 5.5 and the 95.9 percent from GPT-6 Astra.

Without tools, the model achieves a score of 72.8 percent on BenchCAD, nearly matching the 73.0 percent performance of Claude Opus 5.5. Moonshot indicates that the model is particularly strong in parametric modeling for mechanical parts, architectural models, and configurable products. It can write and run scripts for Blender and CAD software without requiring Model Context Protocol connectors. To support this, the company has launched a feature called 3D Make on Kimi.com, allowing users to preview, edit, and export three-dimensional models directly.

Despite the performance upgrades, the structural dimensions of the model remain consistent with its predecessor. Kimi K3.1 Pro retains a total of 2.8 trillion parameters, with 104 billion active parameters. The context window spans one million tokens, and the model processes both text and image inputs. Artificial Analysis notes that the model sits at the top of the Pareto frontier, meaning no model with a comparable parameter count scores higher on the Intelligence Index. This efficiency allows Moonshot to reduce API prices by 33 percent. Input costs have dropped to $2 per million tokens from $3, output costs to $10 from $15, and cache hits to $0.20 from $0.30.

The cost per Intelligence Index task is $1.98, which is nearly identical to the $2.00 cost for the previous model. This similarity occurs because the new model generates 14 percent more output tokens per task. In the realm of cybersecurity, Moonshot has made deliberate choices regarding offensive capabilities. The company observed that vulnerability research skills improved as post-training scaled, due to cyber-related content in the training corpus. For the open-weights release, Moonshot removed cyber-related data from both training phases. On the internal xdev-bench-hard test, which reproduces recent software vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After data cleaning, K3.1 Pro solved 81.0 percent. DeepSeek-V4.1-Flash reached 85.7 percent on the same test, while the older Kimi K3 scored 33.3 percent. Moonshot argues that while open-weights models should not dominate the absolute frontier of offensive cyber capabilities, they should gradually improve to support defensive tasks such as auditing legacy systems.