Moonshot AI Narrows Gap With Western Leaders via Kimi K3.1 Pro

Moonshot AI has released Kimi K3.1 Pro, a post-trained reasoning model that immediately secured the fourth position on the Artificial Analysis Intelligence Index. The release marks a significant milestone for the Chinese company, as the new model scores 55 points, placing it just three points behind the current leader, Anthropic’s Claude Opus 5.5. This achievement highlights the rapid convergence of Chinese and Western artificial intelligence capabilities, challenging the long-standing dominance of proprietary models in high-stakes technical benchmarks.

The most substantial gains in K3.1 Pro are evident in agentic tasks, where the model demonstrates improved autonomy and reliability. On the GDPval-AA benchmark, the new iteration reaches an Elo rating of 1,846, a 312-point improvement over its predecessor, Kimi K3. This performance places it narrowly behind Claude Opus 5.5 and Muse Wonder Full. Furthermore, on Terminal-Bench 4.0, Moonshot reports a success rate of 65.7 percent, a stark contrast to the 13 percent recorded by the previous version. This leap in terminal execution capability suggests the model is better suited for complex software development workflows that require multi-step reasoning and tool usage.

Beyond coding, the model exhibits marked improvements in 3D modeling and spatial reasoning. K3.1 Pro can interact directly with major game engines such as Unity, Unreal Engine, and Godot, handling both minor adjustments and full project construction. On the BenchCAD benchmark, which tests computer-aided design capabilities, the model scores 92.7 percent when utilizing external tools, closely trailing Claude Opus 5.5. Without external assistance, it achieves 72.8 percent, nearly matching the performance of its primary competitor. Moonshot indicates that the model is particularly effective at parametric modeling of mechanical parts and architectural structures, a feature that could accelerate workflows in engineering and design sectors.

Economic efficiency represents another critical aspect of this release. While the model maintains the same architectural dimensions as its predecessor, with 2.8 trillion total parameters and 104 billion active ones, the cost of access has decreased significantly. API pricing has dropped by 33 percent, with input costs now standing at $2 per million tokens and output costs at $10 per million tokens. This reduction in price, combined with the model’s position at the top of the Pareto frontier for its size, makes it a compelling option for developers seeking high-performance reasoning without the premium price tag typically associated with frontier models.

Moonshot has also addressed concerns regarding model hallucinations, a common issue in large language models. The company reports that claim-level hallucinations on its internal Kimi Hallu Bench Hard dropped from 23.4 percent in K3 to 8.3 percent in K3.1 Pro. While this figure remains higher than GPT-6 Astra’s 5.5 percent, the company notes that this benchmark serves as a stress test rather than a measure of general error rates. Additionally, the model has shown improvements in financial applications, gaining 7.2 points on Financial Modeling and 11.6 points on Financial Research benchmarks, indicating enhanced utility for professional office tasks.

A notable strategic decision in the development of K3.1 Pro involves its cybersecurity capabilities. Moonshot observed that vulnerability research abilities increased as post-training compute scaled, due to cyber-related content in the training corpus. To manage this for the upcoming open-weights release, planned for November 4, 2026, the company stripped cyber-related data from both training phases. This adjustment reduced the model’s performance on the xdev-bench-hard test from 90.5 percent to 81.0 percent. The company aims to balance offensive capabilities with defensive needs, arguing that open models should not push the absolute frontier of offensive cyber capabilities while still serving important auditing functions.

The release of K3.1 Pro also includes a high-speed variant designed for faster inference. Moonshot states that K3.1 Pro Highspeed runs two to three times faster than the standard version, offering a trade-off between latency and peak performance. Both models are currently available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. If the open-weights release proceeds as scheduled, K3.1 Pro would become the highest-scoring open model, shrinking the performance gap between open and proprietary systems from 12 points to just three.

The competitive landscape for AI models continues to tighten, with Chinese developers closing in on the performance of established Western firms. The combination of improved agentic reliability, lower costs, and specific optimizations for 3D and coding tasks positions Kimi K3.1 Pro as a serious contender for enterprise adoption. As the gap between open and proprietary models narrows, the market may see further pressure on pricing and performance, benefiting developers and end-users who rely on these advanced reasoning systems.