Moonshot AI Narrows Gap With Western Leaders in New Model Release

A significant shift in the artificial intelligence landscape has occurred as a leading Chinese developer released a new reasoning model that nearly matches the performance of top-tier Western competitors.

Moonshot AI has unveiled Kimi K3.1 Pro, a post-trained iteration of its previous Kimi K3 architecture. The new model has surged to fourth place on the Artificial Analysis Intelligence Index, securing a score of 55 points. This represents an eleven-point increase over its predecessor and places it only three points behind the current leaders: Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. Should Moonshot adhere to its plan of releasing the model weights by November 4, 2026, Kimi K3.1 Pro would become the highest-scoring open-weights model available. This development would narrow the performance gap between open and proprietary systems from twelve points to a mere three, provided no other major updates are released in the interim. The company is also shipping a variant named K3.1 Pro Highspeed, which operates two to three times faster, with both versions accessible via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

The most substantial improvements in the new model are evident in agentic tasks and coding capabilities. On the GDPval-AA benchmark, the model achieves an Elo rating of 1,846, a 312-point leap from the previous version. It now trails only Claude Opus 5.5 and Muse Wonder Full. In terminal-based evaluations, Moonshot reports a score of 65.7 percent on Terminal-Bench 4.0, surpassing GPT-6 Astra and approaching the performance of Claude Opus 5.5. The predecessor model had managed only 13 percent on this specific test. Additionally, internal testing indicates a sharp reduction in hallucination rates, dropping from 23.4 percent to 8.3 percent on a stress-test benchmark. While GPT-6 Astra maintains a lower error rate of 5.5 percent, the improvement in Kimi is notable. The model also shows gains in finance, with specific improvements in financial modeling and research, and currently tops all tested models on the OfficeQA Pro v2 benchmark for general office tasks.

Beyond standard coding, the model demonstrates enhanced capabilities in three-dimensional modeling. It can interact directly with game engines such as Unity, Unreal Engine, and Godot to handle both minor fixes and full project builds. On the BenchCAD benchmark, the model scores 92.7 percent when using external tools, closely following Claude Opus 5.5 and GPT-6 Astra. Without tools, its score of 72.8 percent nearly matches that of Claude Opus 5.5. Moonshot highlights its strength in parametric modeling for mechanical parts and architecture, noting that it can write and run scripts for Blender and CAD software without requiring specific connectors. To support these features, the company launched a new tool called 3D Make on its website, allowing users to preview, edit, and export models.

The architecture of Kimi K3.1 Pro retains the same dimensions as its predecessor, featuring 2.8 trillion total parameters with 104 billion active ones. It supports a context window of one million tokens and processes both text and images. According to Artificial Analysis, the model sits at the top of the Pareto frontier, meaning no model with a comparable parameter count scores higher on the Intelligence Index. To make the advanced performance more accessible, Moonshot has reduced API pricing by 33 percent. Input costs have dropped to $2 per million tokens, output costs to $10, and cache hits to $0.20. The cost per Intelligence Index task remains stable at $1.98, nearly identical to the previous version, as the model generates 14 percent more output tokens.

Moonshot has also addressed cybersecurity considerations in its technical documentation. The company observed that vulnerability research capabilities grew as post-training scaled, due to cyber-related content in the training corpus. For the open-weights release, the team removed cyber-related data from both training phases. On an internal benchmark that reproduces recent vulnerabilities in software like the Linux kernel, the uncleaned model solved 90.5 percent of tasks. After data cleaning, the score settled at 81.0 percent. This compares to 85.7 percent for DeepSeek-V4.1-Flash and 33.3 percent for the previous Kimi K3. The company stated that open-weights models should not push the absolute frontier of offensive cyber capabilities, though they aim to gradually increase their defensive utility, such as auditing legacy systems. This stance follows earlier findings by Anthropic suggesting that even mid-tier models can write exploits at levels comparable to top proprietary systems.