55 Points: How a Chinese AI Model Just Tied for Third Place
Moonshot AI's new release has surged up the global rankings, significantly narrowing the performance gap between open and proprietary systems.
By Allison Rojas ·
The number 55 has become a pivotal metric in the current AI race. That is the score achieved by Kimi K3.1 Pro on the Artificial Analysis Intelligence Index, a result that places the model in fourth position globally. This specific score signals a major shift in the competitive landscape, as it sits just three points behind the leading proprietary systems. The achievement marks a significant leap for Chinese developer Moonshot AI and its latest reasoning model.
This new release represents a post-trained evolution of the previous Kimi K3 architecture. The company invested significantly more compute resources into the post-training phase to drive these improvements. The effort paid off with an eleven-point increase over its predecessor. As a result, only three models currently hold a higher rank. These are Claude Opus 5.5 with 58 points, Claude Sonnet 5.5 with 56 points, and Muse Wonder Full also with 56 points.
The trajectory of this model could reshape the open-weights sector if the release schedule holds. Moonshot plans to make the model weights available by November 4, 2026. If that happens, K3.1 Pro would become the highest-scoring open model on the market. This move would reduce the performance gap between the best open and proprietary models from twelve points down to just three. Such a reduction would fundamentally alter the value proposition for organizations choosing between open and closed systems.
Performance gains are most visible in agentic tasks, where the model shows its strongest capabilities. On the GDPval-AA benchmark, the system achieves 1,846 Elo points. This represents a 312-point improvement over the older K3 version. The score places it just behind Claude Opus 5.5 and Muse Wonder Full. In coding tasks, the results are equally striking, with the model reporting a 65.7 percent success rate on Terminal-Bench 4.0. This far surpasses the previous version, which managed only 13 percent on the same test.
Reliability has also improved, with a sharp drop in hallucination rates. According to internal testing, claim-level hallucinations fell from 23.4 percent to 8.3 percent. While the company notes this is a stress test rather than a general error measure, it indicates a more stable output. The model also performs well in specialized fields like finance, gaining points in both modeling and research tasks. It currently tops all tested models on the OfficeQA Pro v2 benchmark for general office tasks.
The model demonstrates specific strengths in 3D modeling and interaction with game engines. It can handle tasks in Unity, Unreal Engine, and Godot, ranging from quick fixes to full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools, a figure close to the top proprietary competitors. Without tools, it still achieves 72.8 percent, nearly matching the performance of the leading models. Moonshot has also launched a feature called 3D Make, which allows users to preview, edit, and export 3D models directly through the platform.
Despite the performance boost, the underlying architecture remains consistent in size. The model retains 2.8 trillion total parameters with 104 billion active ones. It supports a context window of one million tokens and handles both text and images. However, the cost of using the API has decreased by 33 percent. Input costs now stand at $2 per million tokens, while output costs are $10 per million tokens. This pricing strategy positions the model competitively against other high-performance options.
Moonshot has also addressed cybersecurity concerns in its technical documentation. The company observed that vulnerability research capabilities grew as post-training scaled up. To prepare for the open-weights release, they removed cyber-related data from both training phases. This adjustment lowered the model's score on an internal cybersecurity benchmark from 90.5 percent to 81.0 percent. The company believes this balance is necessary, as open models should not push the absolute frontier of offensive capabilities while still serving defensive purposes like auditing legacy systems.
The release includes a high-speed variant called K3.1 Pro Highspeed, which runs two to three times faster than the standard version. Both versions are currently available through various Moonshot platforms. The market response to these changes will be closely watched, particularly as the November deadline for the weight release approaches. The industry is now focused on whether this new standard will prompt further adjustments from other major AI providers in the coming months.
