Moonshot AI Narrows Gap With Claude Opus 5.5 via New Kimi K3.1 Pro

Moonshot AI has launched Kimi K3.1 Pro, a post-trained reasoning model that has rapidly climbed to the fourth position on the Artificial Analysis Intelligence Index. The release marks a significant shift in the competitive landscape, as the model scores 55 points, placing it just three points behind the leading proprietary systems. This performance leap represents an 11-point improvement over its predecessor, Kimi K3, signaling that Chinese AI developers are closing the distance to top-tier global competitors with increasing speed and precision.

The new model primarily outperforms its predecessor in agentic tasks, where it demonstrates substantial gains in capability. On the GDPval-AA benchmark, Kimi K3.1 Pro achieves an Elo rating of 1,846, a jump of 312 points from the previous version. This places it slightly behind Claude Opus 5.5 and Muse Wonder Full. In coding environments, the model also shows marked improvement, scoring 65.7 percent on Terminal-Bench 4.0. This result surpasses GPT-6 Astra and comes within a fraction of a point of Claude Opus 5.5, which recorded 66.4 percent. By comparison, the earlier Kimi K3 model managed only 13 percent on that same test.

Beyond raw coding ability, the company reports a sharp reduction in hallucination rates. On its internal Kimi Hallu Bench Hard, claim-level errors dropped from 23.4 percent in the previous model to 8.3 percent in the new version. Moonshot describes this metric as a stress test rather than a measure of general error frequency. While GPT-6 Astra maintains a lower error rate of 5.5 percent on the same benchmark, the improvement for Kimi is notable for reducing inconsistencies in generated content. The model also posted gains in specialized finance tasks, increasing its scores on financial modeling and research benchmarks.

A distinctive feature of Kimi K3.1 Pro is its enhanced capacity for 3D modeling and direct integration with game engines. The model can interact with Unity, Unreal Engine, and Godot to handle both minor fixes and full project builds. On the BenchCAD benchmark, it achieved a score of 92.7 percent when using tools, a result that closely rivals the top proprietary models. Without external tools, the score was 72.8 percent, nearly matching the 73.0 percent recorded by Claude Opus 5.5. Moonshot indicates that the model excels at parametric modeling for mechanical parts and architectural designs, allowing it to write and run scripts for Blender and CAD software without needing additional connectors.

Despite the performance upgrades, Moonshot has kept the underlying architecture dimensions unchanged. The model retains 2.8 trillion total parameters with 104 billion active ones, and it supports a context window of one million tokens. It handles both text and image inputs. According to Artificial Analysis, the model sits on the Pareto frontier, meaning no other model with a comparable parameter count scores higher on the Intelligence Index. This efficiency suggests that the gains come from improved training rather than increased size.

The release also brings a significant cost reduction for developers. API prices have dropped by 33 percent across the board. Input costs are now $2 per million tokens, down from $3, while output costs have fallen to $10 from $15. Cache hits are priced at $0.20, a decrease from $0.30. While the per-task cost on the Intelligence Index is $1.98, which is nearly identical to the previous $2.00, this is because the new model generates 14 percent more output tokens. A faster variant called K3.1 Pro Highspeed is also available, running two to three times quicker than the standard version.

Moonshot has taken specific steps to manage the model's cybersecurity capabilities before its open-weights release. The company observed that vulnerability research skills grew as post-training compute increased, likely due to cyber-related content in the training data. For the upcoming open-weights version, scheduled for release by November 4, 2026, Moonshot removed cyber-related data from both training phases. On an internal test involving known vulnerabilities in software like the Linux kernel, the uncleaned model solved 90.5 percent of tasks, while the cleaned version solved 81.0 percent. This is a significant improvement over the previous model, which solved only 33.3 percent.

The company aims to balance offensive and defensive capabilities in its open-weights strategy. Moonshot believes that open models should not push the absolute frontier of offensive cyber capabilities but should gradually increase in ability to support defensive tasks such as auditing legacy systems. This approach contrasts with some proprietary models that may hide certain capabilities. If the November release proceeds as planned, Kimi K3.1 Pro would become the highest-scoring open-weights model available. This would shrink the gap between open and proprietary models from 12 points to just three, assuming no other major releases occur in the interim.

The launch positions Moonshot as a key player in the current AI race. By maintaining the same parameter count while improving performance and lowering prices, the company has offered a compelling alternative to existing top-tier models. The availability of the model through Kimi.ai, Kimi Work, Kimi Code, and the API means developers can access these improvements immediately. The coming months will be critical as the open-weights version arrives, potentially reshaping how researchers and businesses approach model selection and deployment.