Moonshot's Kimi K3.1 Pro Narrows Gap With Top US AI Models

The competitive distance between Chinese and American frontier AI models has shrunk dramatically following the launch of Kimi K3.1 Pro. Released by Moonshot AI, the new system challenges the dominance of US-based competitors by securing the fourth position on the Artificial Analysis Intelligence Index. With a score of 55, it trails the leading proprietary models by a narrow margin, signaling a tightening race in the high-end reasoning sector.

The performance leap is substantial compared to its predecessor. Kimi K3.1 Pro represents a post-trained evolution of Kimi K3, developed with significantly increased computational resources during the post-training phase. This investment results in an 11-point increase in its index score. Currently, only three models rank higher: Claude Opus 5.5 at 58 points, followed by Claude Sonnet 5.5 and Muse Wonder Full, both at 56 points. This positioning places the new Chinese model squarely in the upper echelon of global AI capabilities.

Strategic timing could further alter the market landscape. Moonshot has committed to releasing the model weights by November 4, 2026. If executed as planned, Kimi K3.1 Pro would become the highest-scoring open-weights model available. This move would reduce the performance gap between open and proprietary systems from 12 points to just three, assuming no other major releases occur in the interim. Such a shift would intensify pressure on closed-source providers to differentiate their offerings beyond raw benchmark scores.

The model’s most significant advancements are evident in agentic tasks and coding. On the GDPval-AA benchmark, the system achieves an Elo rating of 1,846, surpassing its predecessor by 312 points. It narrowly trails Claude Opus 5.5 and Muse Wonder Full in this category. In terminal-based coding tests, Moonshot reports a success rate of 65.7 percent on Terminal-Bench 4.0. This figure exceeds GPT-6 Astra’s 58.2 percent and sits barely behind Claude Opus 5.5’s 66.4 percent, marking a massive improvement from the 13 percent recorded by the original K3.

Reliability has also seen marked improvement, particularly in reducing hallucinations. Internal testing on the Kimi Hallu Bench Hard shows claim-level hallucinations dropping from 23.4 percent in the previous version to 8.3 percent. While Moonshot notes this is a stress test rather than a general error metric, the reduction is significant. However, GPT-6 Astra maintains an advantage with a rate of 5.5 percent on the same benchmark. The new model also posts gains in financial modeling and research, and tops tested models on office task evaluations.

Capabilities in 3D modeling and game development have expanded as well. The system now interacts directly with engines like Unity, Unreal Engine, and Godot, handling both minor fixes and full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools, approaching the performance of top US rivals. Without external tools, it achieves 72.8 percent, nearly matching Claude Opus 5.5. Moonshot highlights its strength in parametric modeling for mechanical parts and architecture, alongside a new feature on Kimi.com that allows users to preview and export 3D designs.

Despite the performance gains, the infrastructure remains unchanged in scale but improved in efficiency. The model retains 2.8 trillion total parameters with 104 billion active ones and supports a one-million-token context window for text and images. It sits at the top of the Pareto frontier for models of its size. More notably, API prices have dropped by 33 percent. Input costs are now $2 per million tokens, output is $10, and cache hits cost $0.20. These reductions make the high-performance model more accessible to developers and enterprises.

Moonshot has also addressed cybersecurity concerns inherent in large language models. The company observed that vulnerability research capabilities grew as post-training scaled, due to cyber-related content in the training data. To manage this, Moonshot removed specific cyber-related data from both training phases for the open-weights release. This cleaning reduced the model’s ability to solve tasks on an internal vulnerability benchmark from 90.5 percent to 81.0 percent. While this is still higher than the previous version’s 33.3 percent, it is lower than DeepSeek-V4.1-Flash’s 85.7 percent, reflecting a deliberate balance between offensive capability and defensive utility.

The release includes a high-speed variant, K3.1 Pro Highspeed, which operates two to three times faster than the standard version. Both variants are currently available through various Moonshot platforms and APIs. As the gap between open and closed models continues to narrow, the industry watches closely to see if this release forces a reevaluation of proprietary AI pricing and performance standards.