New AI model from China cuts costs and boosts performance

For developers and businesses using artificial intelligence tools, the cost of running complex tasks is going down while the quality goes up. A new model released by Chinese company Moonshot AI, called Kimi K3.1 Pro, is priced 33 percent lower than its predecessor. This price drop makes high-performance AI more accessible without sacrificing the capability to handle difficult workloads.

The numbers behind this release are substantial. Kimi K3.1 Pro scored 55 points on the Artificial Analysis Intelligence Index, an 11-point increase over the previous version. This places it in fourth position globally, just three points behind the top proprietary models from Anthropic and Muse. The model also demonstrates significant improvements in specific areas, including agentic tasks, coding, and 3D modeling.

In terms of speed and efficiency, Moonshot is also shipping a variant known as K3.1 Pro Highspeed. According to the company, this version runs two to three times faster than the standard model. Both versions are currently available through various platforms, including Kimi.ai and the Kimi API. The standard model maintains a large capacity, with a context window of one million tokens and the ability to process both text and images.

The model performs particularly well in professional and technical environments. On benchmarks for agentic tasks, it outperforms several major competitors, including GPT-6 Astra. Moonshot reports that the model has drastically reduced hallucination rates in internal tests, dropping from 23.4 percent to 8.3 percent. It also shows strong gains in financial modeling and research, and it can directly interact with game engines like Unity and Unreal Engine to handle project builds.

A key feature of this release is the plan to make the model weights open to the public by November 4, 2026. If this happens, Kimi K3.1 Pro would become the highest-scoring open-weights model available. This move would narrow the performance gap between open-source and proprietary AI models from 12 points down to just three, assuming other models remain static.

Moonshot has also addressed cybersecurity concerns regarding the new model. The company notes that vulnerability research capabilities tend to grow with more post-training. To manage this, they removed cyber-related data from the training phases for the open-weights version. While the uncleaned version solved 90.5 percent of specific security tasks, the cleaned version solved 81.0 percent. Moonshot aims to balance offensive capabilities with defensive uses, such as auditing legacy systems, while acknowledging that even smaller models can write exploits.