Moonshot AI’s Kimi K3.1 Pro Narrows Gap With Anthropic

Moonshot AI has officially launched Kimi K3.1 Pro, a new reasoning model that has quickly climbed the rankings to fourth place on the Artificial Analysis Intelligence Index. The system scores 55 points, placing it just three points behind Anthropic’s Claude Opus 5.5, which leads the proprietary field. This release marks a significant shift in the competitive landscape, as the Chinese firm’s latest offering is now within striking distance of the top-tier Western models that have historically held the lead in overall intelligence benchmarks.

To understand the significance of this leap, it helps to look at the recent trajectory of Moonshot’s models. The predecessor, Kimi K3, was a solid performer but sat considerably lower on the index. By applying significantly more post-training compute, the company managed to boost the score by 11 points in a single iteration. Currently, only three models rank higher than K3.1 Pro: Claude Opus 5.5 at 58 points, and both Claude Sonnet 5.5 and Muse Wonder Full at 56 points each. If Moonshot follows through on its plan to release the model weights by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available. This move would shrink the gap between open and proprietary systems from 12 points down to just three, assuming no other major releases occur in the interim.

The improvements are most visible in agentic tasks, where the model demonstrates substantial gains over its predecessor. On the GDPval-AA benchmark, K3.1 Pro reaches 1,846 Elo, a jump of 312 points over K3. This places it just behind Claude Opus 5.5 and Muse Wonder Full. In terminal-based tasks, the results are even more striking. On Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent, surpassing GPT-6 Astra and coming within 0.7 percent of Claude Opus 5.5. By comparison, the older K3 model managed only 13 percent on that same test. The model also shows a sharp reduction in hallucinations. On Moonshot’s internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent to 8.3 percent, though GPT-6 Astra still holds an edge at 5.5 percent.

Beyond general reasoning, the model excels in specific professional domains. On internal finance benchmarks, K3.1 Pro gained 7.2 points in Financial Modeling and 11.6 points in Financial Research. For general office work, the company claims the model tops all tested competitors on OfficeQA Pro v2. The model also integrates directly with game engines like Unity, Unreal Engine, and Godot, handling everything from quick fixes to full project builds. In 3D modeling, it scores 92.7 percent on the BenchCAD benchmark when using tools, close to the top performers. Moonshot notes it is particularly strong at parametric modeling of mechanical and architectural parts, and it can write scripts for Blender and CAD software without needing external connectors.

Despite the performance gains, the model retains the same architectural footprint as its predecessor. It maintains 2.8 trillion total parameters with 104 billion active ones, and supports a context window of one million tokens for both text and images. According to Artificial Analysis, it sits at the top of the Pareto frontier, meaning no model with a comparable parameter count scores higher on the Intelligence Index. For developers and businesses, the cost structure is also more favorable. API prices have dropped by 33 percent. Input costs are now $2 per million tokens, down from $3, while output costs are $10, down from $15. Cache hits cost $0.20, down from $0.30. Because the model generates 14 percent more output tokens, the cost per Intelligence Index task remains nearly identical to the previous version at $1.98.

Moonshot has also addressed cybersecurity capabilities in its technical release. The company noted that vulnerability research abilities grew as post-training scaled up, due to cyber-related content in the training data. For the open-weights version, they stripped this data from both training phases to balance offensive and defensive capabilities. On an internal test called xdev-bench-hard, which reproduces recently disclosed vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After data cleaning, K3.1 Pro solves 81.0 percent. This is lower than DeepSeek-V4.1-Flash, which reaches 85.7 percent, but significantly higher than the previous Kimi K3, which scored 33.3 percent. The company argues that while no open model should dominate the offensive frontier, they should gradually improve to support defensive tasks like auditing legacy systems.

The launch includes a high-speed variant called K3.1 Pro Highspeed, which runs two to three times faster than the standard version. Both models are available immediately through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Users can also access a new feature called 3D Make on Kimi.com, which allows for the preview, editing, and export of 3D models. This release signals that the gap between open-source and proprietary AI is narrowing, with Chinese firms now competing directly on both price and performance at the top end of the market.

Looking ahead, the next major milestone is the scheduled release of the model weights by early November 2026. If this timeline holds, developers will have access to the highest-scoring open-weights model on the index. This could accelerate adoption in environments where data privacy or on-premise deployment is critical, as the performance gap with proprietary leaders becomes negligible. The market will be watching closely to see if this release triggers a broader shift in how enterprises choose their foundational AI models, potentially favoring open weights that offer similar intelligence at a lower cost and greater flexibility.