Moonshot AI cuts prices for new Kimi model

Developers and businesses using artificial intelligence tools can now access a more affordable option from a major Chinese provider. Moonshot AI has lowered the price of its latest large language model by one third while keeping the underlying architecture unchanged. This move aims to make advanced reasoning capabilities more accessible to a wider range of users without requiring significant changes to existing infrastructure.

The new release, named Kimi K3.1 Pro, represents a significant upgrade over its predecessor. It achieved a score of 55 points on the Artificial Analysis Intelligence Index. This result places it in fourth position globally, just behind Anthropic's Claude Opus 5.5 and two other top-tier proprietary models. The jump from the previous version marks an 11 point increase in overall performance.

The model demonstrates particular strength in agentic tasks and coding. On the GDPval-AA benchmark, it recorded an Elo score of 1,846. This figure is only slightly lower than the leading competitors. In terminal-based coding tests, the system achieved a 65.7 percent success rate. This performance surpasses several other major models and shows a massive improvement over the initial K3 version, which scored only 13 percent on the same test.

Beyond standard text processing, the software handles 3D modeling and game engine integration. It can interact directly with platforms like Unity and Unreal Engine. In CAD benchmarks, the model achieved high accuracy both with and without external tools. Moonshot also introduced a new feature that allows users to preview and export 3D models directly through the web interface. This addition supports workflows for mechanical and architectural design.

To manage costs, Moonshot maintained the model's 2.8 trillion total parameters while reducing input and output token prices. Input costs are now $2 per million tokens, down from the previous $3. Output prices dropped to $10 from $15. The company plans to release the model weights by early November. If this timeline holds, the system would become the highest scoring open weights model available, narrowing the gap with proprietary rivals.

The provider also addressed cybersecurity considerations in its technical documentation. It noted that vulnerability research capabilities grew as post-training scaled. To balance offensive and defensive uses, the team removed specific cyber-related data from the training set for the open release. This adjustment resulted in a lower score on internal security tests compared to the uncleaned version. However, the company believes this approach is appropriate for open models that often support defensive auditing tasks.

Both the standard version and a faster variant called K3.1 Pro Highspeed are available now. The high-speed option runs two to three times quicker according to the company. These tools are accessible through various platforms including Kimi.ai and the developer API. The release continues the trend of Chinese AI firms competing directly with top US technology companies on both performance and price.