China's Kimi K3.1 Pro closes in on top US models, independent test finds

Fourth. That is the rank Kimi K3.1 Pro, the new model from Chinese start-up Moonshot AI, earned on the index run by Artificial Analysis, an independent organization that evaluates AI models. Only three closed American models rank ahead of it: Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5, and Meta's Muse Wonder Full.

Moonshot unveiled the model on October 23. It is available now through its paid cloud services. The company plans to release the weights, meaning the model itself, which anyone can run on their own servers, by November 4, according to SiliconANGLE. At that point, according to Artificial Analysis, it would become the best open model in its ranking.

What does the independent test measure?

Kimi K3.1 Pro scores 55 points on the Artificial Analysis index, against 58 for Claude Opus 5.5 and 56 for Claude Sonnet 5.5 and Muse Wonder Full. Its predecessor Kimi K3, released in July, topped out at 44. The clearest progress is in agentic tasks, where the AI chains actions on its own: on Terminal-Bench 4.0, which measures command-line work, its success rate jumps from 13% to 67%, ahead of Sonnet 5.5 (64%). On the AA-Omniscience index, which assesses knowledge and reliability, its hallucination rate falls to 44%.

The figures published by Moonshot itself go further. On a complex financial analysis test, OfficeQA Pro v2, the company says it outperforms OpenAI's GPT-6 Astra and Claude Opus 5.5. These measurements have not been independently verified.

Why the model lags in cybersecurity

The most unexpected detail concerns computer security. According to SiliconANGLE, Moonshot deliberately removed cybersecurity-related information from the data used to fine-tune the model, to address security risks. As a result, on standard cybersecurity tests, Kimi K3.1 Pro remains well behind GPT-6 Astra and Opus 5.5. The company says it wants to add more advanced vulnerability analysis capabilities in future models.

It remains to be seen what those who download the weights will do with this choice: an open model can be retrained by anyone, including to restore what was removed.

What this means for a European company

First, the price: $2 per million tokens (the units of text processed) for input and $10 for output, a third less than Kimi K3. But the model writes more: an average of 55,000 tokens produced per task, against 48,000 for K3. As a result, the index's cost per task, $1.98, stays roughly the same as for its predecessor, Artificial Analysis calculates.

Second, releasing the weights would make it possible to run the model on European infrastructure, without sending data to a Chinese provider. But with 2.8 trillion parameters in total, of which 104 billion are active at each step according to Artificial Analysis, it will take a substantial fleet of graphics cards. The license is not yet known. Kimi K3 was released under a Moonshot-specific license rather than a standard open-source one, according to the specialist press: read it carefully before any commercial use.

The gap between open and closed models is narrowing: it would shrink from 12 to 3 points on the Artificial Analysis index. In mid-October, Meta's best open model was already trailing two Chinese rivals. See you on November 4 to check that Moonshot keeps its promise.