Moonshot AI releases Kimi K3.1 Pro to rival Anthropic's Claude

Closing the gap with Claude

Moonshot AI recently launched Kimi K3.1 Pro, its first major upgrade since the K3 release. The company is positioning this new model as a direct competitor to high-end closed models like Anthropic's Claude Opus 5.5. According to benchmarks reported by Startup Fortune, the model achieved a score of 65.7% on Terminal-Bench 4.0.

Terminal-Bench 4.0 is a specific test designed to score coding agents on realistic tasks within resource-constrained environments. On this same test, the model outperformed GPT-6 Astra, which managed a score of 58.2%. This puts K3.1 Pro within striking distance of the top-tier performance usually reserved for proprietary systems.

While these figures come from Moonshot's own internal testing and have not been verified by an independent lab, they represent a significant milestone. The results suggest that open-weight models are reaching a level of sophistication that makes them viable alternatives to paid subscriptions. This shift is particularly relevant as enterprise interest in open-weight models continues to grow.

Architecture and technical specs

The K3.1 Pro model is designed to be enormous, utilizing a mixture-of-experts architecture. This setup allows the model to carry roughly 2.8 trillion total parameters while only using 104 billion active parameters at any single time. This architecture helps the model manage complex tasks without requiring every single parameter to be active for every calculation.

In addition to its parameter count, the model supports a 1 million token context window. This refers to the amount of information the model can process in a single session, which is vital for analyzing long documents or large codebases. Moonshot also released a faster companion variant called Kimi K3.1 Pro Highspeed for users who need quicker responses.

The company attributed the performance gains to heavier post-training rather than simply building a larger base model. This focus on training refinement helped improve multi-step execution and tool use. On an internal test called Kimi Code Bench 3.0 Long-horizon, the model scored 61.2%, a notable jump from the 55.95% scored by its predecessor.

Accuracy and 3D capabilities

One of the most significant improvements in K3.1 Pro involves reducing hallucinations, which are instances where an AI generates false or fabricated information. On the Kimi Hallu Bench Hard stress test, the claim-level hallucination rate dropped from 23.4% in the K3 model to just 8.3% in the K3.1 Pro version. This reduction is a critical factor for enterprise buyers who need reliable outputs for production environments.

Beyond standard text and code, Moonshot has pushed the model toward specialized 3D tools. K3.1 Pro can now work directly inside software like Blender, Unreal Engine, Godot, and CAD programs. It is capable of generating full game projects and parametric 3D models rather than just providing small snippets of code.

In a test called BenchCAD, which measures 3D reconstruction from multi-view renderings via CAD code generation, the model scored 92.7%. This performance is very close to the 96.2% scored by Claude Opus 5.5 and the 95.9% scored by GPT-6 Astra. These results highlight the model's versatility across different creative and engineering fields.

Release strategy and safety

Moonshot AI is using a deliberate two-week window between the current API launch and the release of the full open weights. The company will make the weights free to download on November 4. This strategy allows Moonshot to monetize the hosted version of the model before the public can run it on their own private hardware.

To address security concerns, Moonshot stated it stripped cyber-related data out of both training stages before making the weights public. This move is intended to ensure that the model remains a capable coding tool without becoming a tool for hacking. The company aims to maintain the model's general coding ability while minimizing risks associated with open-source distribution.

The pricing for the model currently sits at $3 to $15 per million input and output tokens. This is a significant increase from the ultra-cheap pricing seen in other Chinese models like DeepSeek. It signals a shift in the market where top Chinese labs are now competing on high-end capability rather than just low costs.

The rise of open weights

The release of Kimi K3.1 Pro occurs during a period of rapid adoption for open-weight models in the corporate world. A survey by Constellation Research and TechStrong.ai found that open-weight models already account for 34% of enterprise AI token usage. This number is expected to rise to 41% within the next year as companies seek more control over their deployments.

Enterprises are moving toward these models because they offer lower costs and more flexibility for internal use. Google's Gemma has already overtaken Meta's Llama as the most-used open-weight family in production. Kimi K3.1 Pro enters this space at a time when U.S. policy is becoming more skeptical of Chinese AI products.

For many companies, the decision to use a model comes down to the balance between performance and ownership. If an open-weight model can perform within a fraction of a point of a closed frontier model, it changes the calculus for IT departments. They can choose to buy GPUs and run Kimi themselves rather than paying for a recurring subscription to a closed service.