Moonshot AI's Kimi K3.1 Pro Narrows the Gap With Top Global Models

Moonshot AI has officially launched Kimi K3.1 Pro, a post-trained iteration of its previous model that immediately secured the fourth position on the Artificial Analysis Intelligence Index. The new release scores 55 points, placing it just three points behind the leading proprietary systems. This performance marks an eleven-point improvement over its predecessor, signaling a substantial leap in the company’s technical capabilities.

The primary drivers of this ranking surge are found in agentic tasks, coding proficiency, and 3D modeling. On the GDPval-AA benchmark, the model achieves an Elo rating of 1,846, surpassing the previous version by more than 300 points. In terminal-based coding evaluations, specifically Terminal-Bench 4.0, K3.1 Pro records a score of 65.7 percent. This result outperforms GPT-6 Astra and trails only Claude Opus 5.5, which scored 66.4 percent on the same test. Notably, the earlier K3 model managed only 13 percent on this benchmark, highlighting the dramatic efficiency gains achieved through additional post-training compute.

Accuracy improvements are equally pronounced, with a sharp reduction in hallucination rates. Internal testing on the Kimi Hallu Bench Hard shows claim-level hallucinations dropping from 23.4 percent to 8.3 percent. While GPT-6 Astra maintains a slight edge with a 5.5 percent error rate on this specific stress test, Moonshot emphasizes that the new model significantly enhances reliability for general office and financial tasks. The model also demonstrates strength in direct interaction with game engines such as Unity and Unreal, capable of handling everything from minor fixes to complete project builds without requiring external connectors for scripts in Blender or CAD software.

Despite these performance upgrades, the structural architecture of the model remains unchanged. K3.1 Pro retains the 2.8 trillion total parameter count and 104 billion active parameters of its predecessor, alongside a one-million token context window. However, the cost of access has decreased by 33 percent. API input prices are now $2 per million tokens, while output costs stand at $10, and cache hits are priced at $0.20. A faster variant, labeled K3.1 Pro Highspeed, is also available, reportedly operating two to three times quicker than the standard version.

A key strategic element of this release is the planned openness of the model’s weights. Moonshot intends to make the weights available by November 4, 2026. If realized, K3.1 Pro would become the highest-scoring open-weights model on the index. This move would reduce the performance disparity between open and proprietary models from 12 points to just three, assuming no other major releases occur in the interim.

The company has also addressed cybersecurity capabilities, noting that offensive potential increased as post-training expanded. To manage this for the open-weights version, Moonshot removed specific cyber-related data from the training process. Internal testing reveals that the cleaned model solves 81.0 percent of tasks on a difficult vulnerability benchmark, compared to 90.5 percent for the uncleaned version. This approach aims to balance defensive utility with controlled offensive potential, a stance that contrasts with some competitors who allow higher scores in this domain.

The launch underscores a tightening race in the AI landscape, where Chinese firms are increasingly challenging the dominance of American proprietary models through aggressive pricing and rapid iteration. The coming weeks will be critical as the open-weights release approaches, potentially reshaping the accessibility and competitive dynamics of high-performance reasoning models.