AI Benchmarks: Moonshot's Kimi K3.1 Pro Narrows Gap With Top Western Models

The race for artificial intelligence supremacy is tightening as a new model from China steps up to challenge the established Western leaders. Moonshot AI has released Kimi K3.1 Pro, a significant upgrade to its previous version that is currently closing the performance gap with top-tier proprietary systems. This development marks a notable shift in the competitive landscape, where open-weights models are beginning to rival the closed systems that have long held the top spots.

The core claim of this release is that Kimi K3.1 Pro has achieved a substantial leap in capability, landing at fourth place on the Artificial Analysis Intelligence Index with a score of 55. This represents an 11-point increase over its predecessor, Kimi K3. It now trails only three models: Claude Opus 5.5, which scores 58, and two others, Claude Sonnet 5.5 and Muse Wonder Full, both at 56. 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 would reduce the performance difference between the best open models and the top proprietary ones from 12 points to just 3, assuming no other major releases occur in the interim.

The most dramatic improvements are seen in agentic tasks, which involve the model acting autonomously to complete complex workflows. On the GDPval-AA benchmark, the new model scores 1,846 Elo, a 312-point jump from the previous version. It now sits just behind Claude Opus 5.5 and Muse Wonder Full. In coding and terminal tasks, the gains are equally stark. On Terminal-Bench 4.0, the model achieved a 65.7 percent success rate, far exceeding the 13 percent recorded by its predecessor. This result places it ahead of GPT-6 Astra and very close to the performance of Claude Opus 5.5, which scored 66.4 percent.

Reliability has also improved, with a sharp drop in hallucination rates. According to Moonshot’s internal stress test, claim-level hallucinations fell from 23.4 percent to 8.3 percent. While GPT-6 Astra maintains a slight edge with a 5.5 percent rate, the reduction in K3.1 Pro is significant. The company notes that this metric is a stress test rather than a general measure of error, but it indicates a more grounded model. Additionally, the model has strengthened in finance, with gains of 7.2 points in financial modeling and 11.6 points in financial research. Moonshot claims it tops all tested models on general office tasks via the OfficeQA Pro v2 benchmark.

A key area of expansion is 3D modeling and game development. The model can now interact directly with engines like Unity, Unreal Engine, and Godot, handling tasks ranging from quick fixes to full project builds. On the BenchCAD benchmark, it scored 92.7 percent when using tools, approaching the scores of GPT-6 Astra and Claude Opus 5.5. Without tools, it reached 72.8 percent, nearly matching the leader. Moonshot highlights its strength in parametric modeling for mechanical and architectural designs, and has launched a "3D Make" feature that allows users to preview and export models directly on its platform.

Despite the performance boosts, the underlying architecture remains the same size as before, with 2.8 trillion total parameters and 104 billion active ones. It supports a one-million-token context window and processes both text and images. The main financial incentive for users is a 33 percent reduction in API prices. Input costs have dropped to $2 per million tokens, and output costs to $10. While the price per token is lower, the cost per task on the Intelligence Index remains nearly identical to the previous version at $1.98, because the new model generates 14 percent more output tokens to complete tasks.

Moonshot has also addressed cybersecurity concerns, particularly for the upcoming open-weights release. The company found that vulnerability research capabilities grew as post-training scaled up. To manage this, they removed cyber-related data from the training process for the open version. In internal testing, the uncleaned model solved 90.5 percent of security tasks, while the cleaned version solved 81.0 percent. This is comparable to DeepSeek-V4.1-Flash, which scored 85.7 percent, and a vast improvement over the previous Kimi K3, which managed only 33.3 percent. Moonshot states that while they aim to gradually increase the offensive capabilities of open models for defensive auditing, they believe no single open model should push the absolute frontier of offensive cyber capabilities.

The release is available now through various platforms including Kimi.ai and the Kimi API. A variant called K3.1 Pro Highspeed is also offered, which runs two to three times faster. The model sits at the top of the Pareto frontier, meaning no other model with a similar parameter count scores higher on the intelligence index. This release suggests that the distinction between open and proprietary AI is becoming less about raw capability and more about specific optimizations and access models.