Moonshot AI Releases Kimi K3.1 Pro, Beats Opus 5.5, GPT Astra 6 On Some Benchmarks

Chinese AI companies are still catching up with the leading AI labs in the United States.

Chinese AI startup Moonshot AI has released Kimi K3.1 Pro, an improved version of its open-weight AI model with around 3 trillion parameters. Parameters are the numbers a model uses to learn and make decisions. The company says its new model performs as well as, or almost as well as, some of the strongest private AI models from Anthropic and OpenAI in tests of coding, completing tasks on its own, and answering questions that require knowledge.

Kimi K3.1 Pro uses a design called mixture-of-experts (MoE). It has 2.8 trillion parameters in total, but only 104 billion are active at a time. It can work with different types of information, such as text and images, without needing separate systems. It can also handle up to 1 million tokens of information at once. Tokens are small pieces of text or other data that AI models process. Moonshot is also releasing a faster version called Kimi K3.1 Pro Highspeed, which works 2–3 times faster than the regular model. Both versions are already available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The company promises to release the model's open weights by November 4, 2026, allowing others to download and use them.

Moonshot has become known for releasing powerful AI models at low prices and making their weights openly available. Its Kimi K2 model, released last year, became one of the strongest open-source AI models available. Later, its Kimi K2.5 update performed better than Claude 4.5 Sonnet on the Artificial Analysis Intelligence Index. This increased the performance gap between Chinese and American open models in China's favor. With K3.1 Pro, Moonshot wants to move even further ahead and compete with the most advanced AI models, which are usually kept private by their developers.

Moonshot is presenting K3.1 Pro as a tool that can handle entire creative and technical projects, rather than just answer questions like a chatbot. The company says the model can write code for complete projects in game-making tools such as Unreal, Unity, and Godot. It can also create 3D models that can be changed using measurements and settings in Blender and CAD software. Moonshot says the model has also become much more dependable for financial work. Compared with the previous version, it reportedly makes up false information 65% less often when working on financial models.

Kimi K3.1 Pro Performance Tests

According to Moonshot's own test results, Kimi K3.1 Pro performs almost as well as the leading AI models from Anthropic and OpenAI. In some tests, it even does better. However, it still scores slightly below the very best models in some areas.

On Terminal-Bench 4.0, K3.1 Pro scored 65.7%. This test checks how well AI agents can complete realistic coding tasks using a computer's command-line terminal, with limited resources and strict checks for each task. K3.1 Pro scored higher than OpenAI's GPT-6 Astra, which received 58.2%. However, it scored slightly lower than Claude Opus 5.5, which came first with 66.4%.

On BenchCAD, K3.1 Pro scored 92.7% when it was allowed to use tools. This test measures how well AI can create 3D objects from pictures showing different angles by writing CAD code. K3.1 Pro's result was close to GPT-6 Astra's 95.9% and Claude Opus 5.5's 96.2%. Moonshot says its model can also handle most of the steps in a creative project by itself, including making 3D objects, putting scenes together, adding animation, and producing the final images.

Moonshot also says K3.1 Pro has achieved the best result so far on OfficeQA Pro v2, a test created within the company to measure how well AI handles office and business tasks. The model scored 64.4. Compared with the original K3, Moonshot says K3.1 Pro scored 7.2 points higher on financial modeling and 11.6 points higher on financial research. It also scored about 4.2 times as high on tests that imitate real business operations. Examples shared by the company include a complete financial model for NVDA that uses discounted cash flow (DCF) and leveraged buyout (LBO) methods, with connected calculations and formulas that can be checked. Other examples include a presentation for an investment committee about Chinese battery maker CATL and a report on the quantum computing industry that includes citations pointing to specific sources.

Kimi K3.1 Pro Pricing

Kimi K3.1 Pro costs two-thirds as much as the original K3. This continues Moonshot's approach of charging less than leading Western AI companies for access through its API while trying to offer similar performance.

For comparison, the regular K3.1 Pro costs about half as much as Claude Opus 5.5 for both input and output tokens. Claude Opus 5.5 charges $4 per million input tokens and $20 per million output tokens. K3.1 Pro also costs much less than GPT-6 Astra, which charges $10 per million input tokens and $50 per million output tokens. Even the faster K3.1 Pro Highspeed version, which costs $4 per million input tokens and $20 per million output tokens, has the same prices as Claude Opus 5.5. Moonshot also promises that its Highspeed version runs 2–3 times faster than the regular K3.1 Pro.

Both versions of Kimi K3.1 Pro are now available through the Kimi API at platform.kimi.ai.