Moonshot AI Narrows Gap With Top US Rivals In New Model Release

Moonshot AI has officially launched Kimi K3.1 Pro, a significant upgrade that pushes the Chinese company closer to the top of the artificial intelligence rankings. The model now sits at number four on the Artificial Analysis Intelligence Index with a score of 55 points. This places it just three points behind the leading proprietary models from Anthropic and other major players. For California tech professionals, this shift signals that the distance between open and closed AI systems is shrinking rapidly.

The new release is a post-trained version of the previous Kimi K3 model. Moonshot applied significantly more computing power to the training process, which resulted in an 11 point jump in overall performance. Only three models currently rank higher: Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. If the company follows through on its plan to release the model weights by November 4, 2026, K3.1 Pro will become the highest scoring open weights model available. That move would reduce the performance gap between open and proprietary models from 12 points to just three.

Users can access K3.1 Pro immediately through Kimi.ai, Kimi Work, Kimi Code, and the standard API. Moonshot is also shipping a variant called K3.1 Pro Highspeed. This version runs two to three times faster than the standard release, offering a trade off between speed and maximum depth. Both options are available for developers and businesses looking to integrate advanced reasoning capabilities into their workflows.

The most dramatic improvements appear in agentic tasks, where the model performs complex, multi step actions. On the GDPval-AA benchmark, K3.1 Pro achieves an Elo rating of 1,846. This is a substantial increase of 312 points over its predecessor. It now trails only Claude Opus 5.5 and Muse Wonder Full in this specific metric. On Terminal-Bench 4.0, the model scores 65.7 percent. This result surpasses GPT-6 Astra and comes within a fraction of a percent of the top performing Claude model. The previous K3 model managed only 13 percent on this same test.

Reliability has also improved significantly. Moonshot reports that claim level hallucinations dropped from 23.4 percent to 8.3 percent on its internal stress test. While GPT-6 Astra still holds an edge with a 5.5 percent rate, the reduction shows a major step forward in accuracy. In financial applications, the model gained 7.2 points in modeling and 11.6 points in research. It also tops all tested models on the OfficeQA Pro v2 benchmark for general office tasks.

Specialized capabilities in coding and 3D modeling represent another key area of growth. The model works directly with game engines such as Unity, Unreal Engine, and Godot. It handles everything from quick bug fixes to full project builds. On the BenchCAD benchmark, K3.1 Pro scores 92.7 percent when using tools. This is close to the scores of Claude Opus 5.5 and GPT-6 Astra. Without external tools, it hits 72.8 percent, nearly matching the top proprietary model. A new feature called 3D Make allows users to preview, edit, and export models directly on the platform.

Despite the performance boost, the underlying architecture remains unchanged. The model retains 2.8 trillion total parameters with 104 billion active ones. The context window spans one million tokens and supports both text and images. According to Artificial Analysis, it sits at the top of the Pareto frontier for its size. No other model with a comparable parameter count scores higher on the intelligence index. However, the cost of using the model has dropped by 33 percent. Input tokens now cost $2 per million, down from $3, and output tokens cost $10, down from $15.

Moonshot has placed a strong emphasis on cybersecurity in its technical documentation. The company observed that vulnerability research capabilities grew as post training scaled up. To manage this for the open weights release, they stripped cyber related data from both training phases. On an internal benchmark that reproduces recent software vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test, while the older Kimi K3 scored only 33.3 percent.

The company believes that open weights models should not push the absolute frontier of offensive cyber capabilities. At the same time, they want to gradually raise the bar because these models also serve defensive purposes, such as auditing legacy systems. Anthropic has previously noted that even lower scoring models can write exploits at levels comparable to top tier systems. This release highlights the ongoing tension between making powerful AI available and managing its potential risks. The next few weeks will be crucial as developers test the new weights and the market reacts to the lower price points.