Moonshot AI releases Kimi K3.1 Pro, narrowing gap with top-tier competitors

Moonshot AI has introduced Kimi K3.1 Pro, a new reasoning model that has rapidly ascended to the fourth position on the Artificial Analysis Intelligence Index. By achieving a score of 55 points, the model has narrowed the distance to the leading proprietary competitors, falling only three points behind Claude Opus 5.5, which currently holds the top spot with a score of 58. This release represents a substantial strategic move for the Chinese AI firm, as it significantly reduces the performance disparity between open-weights and closed-source architectures.

The update is a post-trained iteration of the predecessor Kimi K3, developed with significantly increased computational resources during the training phase. According to Artificial Analysis, the performance jump is marked by an 11-point improvement over the previous version. Only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full currently outperform K3.1 Pro on the index. If Moonshot proceeds with its planned release of the model weights by November 4, 2026, K3.1 Pro will become the highest-scoring open-weights model available, shrinking the overall gap between open and proprietary models from 12 points to just three.

Moonshot AI also launched a variant known as K3.1 Pro Highspeed, which the company states operates two to three times faster than the standard version. Both models are currently accessible through Kimi.ai, Kimi Work, Kimi Code, and the company's API, offering immediate availability to developers and enterprise users seeking enhanced agentic capabilities.

The most pronounced advancements in K3.1 Pro are evident in agentic tasks and coding. On the GDPval-AA benchmark, the model records an Elo rating of 1,846, a 312-point increase over K3. This places it just behind Claude Opus 5.5 and Muse Wonder Full. In terminal-based coding tests, K3.1 Pro achieves a 65.7 percent score on Terminal-Bench 4.0, surpassing GPT-6 Astra and closely trailing Claude Opus 5.5. Notably, the previous version, K3, had managed only 13 percent on the same test, highlighting the magnitude of the recent improvements.

Reliability has also improved, with hallucination rates decreasing sharply. On Moonshot’s internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent in K3 to 8.3 percent in the new model. While GPT-6 Astra maintains a lower error rate at 5.5 percent, Moonshot emphasizes that this benchmark serves as a stress test rather than a measure of general error rates. The model also demonstrated gains in financial applications, scoring 7.2 points higher in financial modeling and 11.6 points higher in financial research compared to its predecessor.

In the realm of 3D modeling and development, K3.1 Pro integrates directly with game engines such as Unity, Unreal Engine, and Godot. It handles both minor adjustments and complete project builds, scoring 92.7 percent on the BenchCAD benchmark when utilizing tools. This performance is close to that of Claude Opus 5.5 and GPT-6 Astra. The model excels in parametric modeling of mechanical parts and architectural structures, capable of writing and executing scripts for Blender and CAD software without requiring MCP connectors. To support these features, Moonshot introduced a "3D Make" tool on Kimi.com for user previews and exports.

Despite the performance enhancements, the model retains the same architectural dimensions as K3, featuring 2.8 trillion total parameters with 104 billion active ones. It supports a context window of one million tokens and processes both text and images. Artificial Analysis notes that K3.1 Pro sits at the top of the Pareto frontier, meaning no other model with a comparable parameter count achieves a higher intelligence score.

Economic efficiency has improved with a 33 percent reduction in API pricing. Input costs have fallen to $2 per million tokens from $3, while output costs dropped to $10 from $15. Cache hit costs are now $0.20 per million tokens, down from $0.30. Although the model generates 14 percent more output tokens, the cost per Intelligence Index task remains nearly identical to the previous version at $1.98.

Moonshot AI dedicated a portion of its technical release to cybersecurity considerations, noting that vulnerability research capabilities grew as post-training scaled. For the upcoming open-weights release, the company removed cyber-related data from both training phases to manage offensive capabilities. On the internal xdev-bench-hard test, which reproduces recent vulnerabilities, the cleaned model solves 81.0 percent of tasks, compared to 90.5 percent before cleaning. This approach aims to balance defensive utility with a cautious limit on offensive cyber potential, a topic of ongoing discussion in the AI industry.

The next critical step for Moonshot AI is the scheduled release of the model weights by November 2026. This move will test the hypothesis that open-weights models can maintain competitive parity with proprietary systems at scale. The Pacific Northwest tech sector, along with global developers, will likely scrutinize the model's real-world performance in agentic workflows and its impact on the pricing dynamics of large language model APIs.