Moonshot AI's Kimi K3.1 Pro Narrows Gap With Anthropic in Latest Benchmark Test

Moonshot AI has released Kimi K3.1 Pro, a model that has immediately climbed to the fourth position on the Artificial Analysis Intelligence Index. With a score of 55, the new system sits just three points behind the leading proprietary models. This rapid ascent marks a significant shift in the competitive landscape, particularly as the company plans to make the model’s weights publicly available later this year. If those plans hold, K3.1 Pro would become the highest-scoring open-weights model on the market, effectively compressing the performance gap between open and closed systems from twelve points to just three.

The current leader is Anthropic’s Claude Opus 5.5, which holds a score of 58. It is followed closely by Claude Sonnet 5.5 and Muse Wonder Full, both of which score 56. Moonshot’s new release represents a post-trained evolution of its predecessor, Kimi K3, which had a score of 44. The jump of eleven points is attributed to significantly increased compute resources during the post-training phase. For now, the model is accessible through Kimi.ai, Kimi Work, Kimi Code, and the company’s API, alongside a high-speed variant that the company claims runs two to three times faster.

The most pronounced improvements in K3.1 Pro are visible in agentic tasks, where the model demonstrates a substantial leap over its predecessor. On the GDPval-AA benchmark, the new model achieves an Elo rating of 1,846, a gain of 312 points compared to K3. This places it narrowly behind Claude Opus 5.5 and Muse Wonder Full. Performance in terminal-based coding also improved drastically; on Terminal-Bench 4.0, K3.1 Pro reported a success rate of 65.7 percent. This figure surpasses GPT-6 Astra and comes within less than one percent of Claude Opus 5.5, whereas the older K3 model managed only 13 percent on the same test.

Beyond raw capability, the model shows better reliability in its outputs. Moonshot reports a sharp decline in hallucinations, with claim-level errors dropping from 23.4 percent in K3 to 8.3 percent in K3.1 Pro on an internal stress test. While this remains higher than the 5.5 percent rate recorded for GPT-6 Astra, the company notes that this benchmark is designed to push the model to its limits rather than measure average error rates. In specialized financial applications, the model also recorded gains of 7.2 points in modeling and 11.6 points in research, while topping all tested models on office task benchmarks.

A notable feature of the release is its enhanced ability to handle 3D modeling and direct integration with game engines such as Unity, Unreal Engine, and Godot. The model can manage everything from minor bug fixes to full project builds. On the BenchCAD benchmark, it scored 92.7 percent when using tools and 72.8 percent without them, figures that closely match the performance of Claude Opus 5.5. Moonshot has also introduced a "3D Make" feature on its website, allowing users to preview, edit, and export three-dimensional models directly through the interface.

Despite the performance gains, the underlying architecture remains consistent with the previous generation. K3.1 Pro retains a total parameter count of 2.8 trillion, with 104 billion active parameters. It supports a one-million-token context window and processes both text and images. According to Artificial Analysis, this places the model at the top of the Pareto frontier, meaning no other model with a similar size achieves a higher intelligence score. The company has also reduced API pricing by 33 percent, lowering input costs to $2 per million tokens and output costs to $10 per million tokens. However, because the model generates 14 percent more tokens, the total cost per task on the intelligence index remains nearly identical to the previous version.

The technical release includes a detailed discussion on cybersecurity, a topic that has become increasingly sensitive as AI capabilities expand. Moonshot noted that vulnerability research skills improved naturally as post-training scaled up, driven by cyber-related content in the general training corpus. To prepare the model for open release, the company removed specific cyber data from both training phases. On an internal benchmark that reproduces known vulnerabilities in software like the Linux kernel, the uncleaned version solved 90.5 percent of tasks. After data cleaning, K3.1 Pro solved 81.0 percent, a figure that still exceeds the 33.3 percent achieved by the older K3 model.

Moonshot’s strategy reflects a balance between defensive utility and offensive potential. The company argues that while no open-weights model should dominate the frontier of offensive capabilities, open models serve a vital defensive role in auditing legacy systems. This approach follows observations by other firms, such as Anthropic, which recently demonstrated that even mid-tier open models can write exploits with near-frontier efficiency. As the industry moves toward November 2026 for the planned release of model weights, the focus will shift to how these open systems perform in real-world security and development environments.

The next few weeks will be critical for validating these benchmark claims in independent tests. If the open-weights release proceeds as scheduled by November 4, the AI community will have a new high-performing tool to evaluate. The primary question remains whether the three-point gap to the proprietary leaders is a temporary plateau or a sign that open-source architectures are finally catching up in efficiency and capability.