AI Benchmark Race: Moonshot's Kimi K3.1 Pro Narrows Gap With Western Leaders

The competitive hierarchy of large language models is shifting as a Chinese entrant closes the distance to the industry's leading proprietary systems. Moonshot AI has launched Kimi K3.1 Pro, a post-trained reasoning model that has secured the fourth position on the Artificial Analysis Intelligence Index with a score of 55. This result places it just three points behind the current market leaders, signaling a significant reduction in the performance gap between open and closed-weight architectures.

The model represents a substantial upgrade from its predecessor, Kimi K3, which scored 44 on the same index. This 11-point increase is attributed to significantly increased post-training compute. Currently, only three models rank higher: Claude Opus 5.5 at 58 points, Claude Sonnet 5.5 at 56 points, and Muse Wonder Full also at 56 points. If Moonshot adheres to its plan to release the model weights by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available, shrinking the disparity between the best open and proprietary models from 12 points to just 3.

Alongside the standard release, Moonshot is deploying a variant named K3.1 Pro Highspeed. The company reports that this version operates two to three times faster than the standard model. Both versions are immediately available through the Kimi.ai platform, Kimi Work, Kimi Code, and the Kimi API, allowing developers and enterprises to access the new capabilities without waiting for the full open-weights release.

The most pronounced improvements in K3.1 Pro are evident in agentic tasks and coding benchmarks. On the GDPval-AA metric, the model achieves 1,846 Elo, a jump of 312 points from K3. This places it closely behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). In terminal-based coding evaluations, specifically Terminal-Bench 4.0, K3.1 Pro records a success rate of 65.7 percent. This surpasses GPT-6 Astra, which scored 58.2 percent, and trails only slightly behind Claude Opus 5.5 at 66.4 percent. Notably, the previous generation K3 managed only 13 percent on this same test.

Reliability has also improved, with hallucination rates dropping sharply. According to Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucinations decreased from 23.4 percent in K3 to 8.3 percent in K3.1 Pro. While GPT-6 Astra maintains an edge at 5.5 percent, Moonshot notes that its benchmark is a stress test rather than a measure of general error rates. The model also shows gains in financial tasks, increasing scores by 7.2 points in Financial Modeling and 11.6 points in Financial Research, while topping all tested models on the OfficeQA Pro v2 benchmark for general office tasks.

Capabilities in 3D modeling and direct game engine integration have been expanded. K3.1 Pro can work directly with Unity, Unreal Engine, and Godot for both minor fixes and full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools, approaching the scores of Claude Opus 5.5 (96.2 percent) and GPT-6 Astra (95.9 percent). Without external tools, it achieves 72.8 percent, nearly matching Claude Opus 5.5's 73.0 percent. Moonshot highlights its strength in parametric modeling of mechanical and architectural parts, and has launched a "3D Make" feature on Kimi.com for user preview and export.

Despite the performance gains, the model architecture remains consistent with K3, retaining 2.8 trillion total parameters with 104 billion active ones. The context window supports one million tokens, handling both text and image inputs. According to Artificial Analysis, the model sits at the top of the Pareto frontier, meaning no other model with a comparable parameter count achieves a higher Intelligence Index score. API prices have been reduced by 33 percent, with input costs dropping to $2 per million tokens and output to $10. The cost per Intelligence Index task remains nearly identical to K3 at $1.98, as the model generates 14 percent more output tokens.

Moonshot has also addressed cybersecurity concerns by modifying its training data for the upcoming open-weights release. The company stripped cyber-related data from both training phases to limit offensive capabilities. On its internal xdev-bench-hard, which tests the reproduction of recently disclosed vulnerabilities in software like the Linux kernel, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent, compared to 85.7 percent for DeepSeek-V4.1-Flash and 33.3 percent for the previous Kimi K3. The company aims to balance offensive potential with defensive utility, such as auditing legacy systems.