China's AI Is Getting Closer to the Very Top

For a long time, the world's most capable AI models have been built by companies in the United States. But that gap is starting to shrink. Chinese AI company Moonshot AI just released its latest model, Kimi K3.1 Pro, and it's making a serious run at the top spot. The new model scores 55 points on the Artificial Analysis Intelligence Index. That puts it in fourth place, just a few points behind Anthropic's Claude Opus 5.5, which sits at 58. It's a significant jump for Moonshot, as its previous model, K3, scored 44 points.

The biggest improvements for K3.1 Pro are in what developers call agentic tasks. This is when an AI can work on a problem step-by-step, using tools to finish a job. For example, on a test called Terminal-Bench 4.0 that measures an AI's ability to use a computer terminal, K3.1 Pro scores 65.7 percent. That's a huge leap from the 13 percent its predecessor managed. It's now performing almost as well as Claude Opus 5.5, which scores 66.4 percent. The model also got much better at coding and working with 3D design software like Blender and CAD programs. On a benchmark called BenchCAD, it scores 92.7 percent with tools, which is close to the top scores from Claude Opus 5.5 and GPT-6 Astra.

Along with these new skills, Moonshot says it has cut down on the model's tendency to make things up. On an internal test for checking false claims, the hallucination rate dropped from 23.4 percent with the old K3 model to 8.3 percent with K3.1 Pro. While another model, GPT-6 Astra, still scores better on that specific test at 5.5 percent, the improvement for Kimi is a clear step forward in making the model more reliable for users who need accurate information.

For people and companies looking to use this technology, there is some good news on the cost. Moonshot says the price to use the model through its API has been cut by 33 percent. The cost for processing input is now $2 per million tokens, down from $3. Output costs have dropped from $15 to $10 per million tokens. This makes the powerful new capabilities more affordable for developers who want to build applications with it.

Perhaps the most important detail for the AI community is what Moonshot plans to do next. The company has stated it will release the model weights by November 4, 2026. Model weights are the core instructions that make an AI model work. By making them public, anyone will be able to download and run K3.1 Pro on their own hardware. If this happens as planned, it would make K3.1 Pro the highest-scoring open-weights model available to the public. This would narrow the performance gap between open and proprietary models from 12 points down to just 3, making powerful AI tools more accessible to everyone.