55 Points: New Chinese AI Model Closes the Gap With Top Proprietary Systems
Moonshot AI's latest reasoning model scores 55 points, placing it just three points behind the industry leaders and narrowing the divide between open and closed systems.
By Owen Turner ·
A new benchmark score has reshaped the competitive landscape in artificial intelligence.
Moonshot AI has released Kimi K3.1 Pro, a reasoning model that now holds the fourth position on the Artificial Analysis Intelligence Index. The system achieved a score of 55, marking an 11 point improvement over its predecessor, Kimi K3. This performance places the new model just three points behind the leading proprietary systems. The top scores currently belong to Anthropic's Claude Opus 5.5 at 58, followed by Claude Sonnet 5.5 and Muse Wonder Full, both at 56.
K3.1 Pro is a post-trained version of K3 that utilized substantially more post-training compute. Moonshot plans to publish the model weights by November 4, 2026. If this release proceeds as scheduled, K3.1 Pro would become the highest-scoring open-weights model available. This development would reduce the gap between open and proprietary models from 12 points to 3, assuming no other changes occur in the interim. Alongside the main model, the company is offering K3.1 Pro Highspeed, a variant designed to run two to three times faster. Both versions are accessible through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
The most significant gains appeared in agentic tasks. On the GDPval-AA benchmark, K3.1 Pro reached 1,846 Elo, which is 312 points above K3. This score trails only Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. In Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent. This figure exceeds GPT-6 Astra at 58.2 percent and sits slightly below Claude Opus 5.5 at 66.4 percent. By comparison, the previous K3 model managed only 13 percent on that specific test.
The company also reports a sharp decrease in hallucinations. On its internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent with K3 to 8.3 percent. Moonshot describes this benchmark as a stress test rather than a measure of general error rates. It notes that GPT-6 Astra still leads this category at 5.5 percent. On internal finance benchmarks, the model gained 7.2 points on Financial Modeling and 11.6 points on Financial Research. Moonshot states that the model tops all tested systems on OfficeQA Pro v2.
Capabilities in 3D modeling and coding have also been highlighted. K3.1 Pro works with game engines including Unity, Unreal Engine, and Godot. It can write and run scripts for Blender and CAD software without requiring MCP connectors. On the BenchCAD benchmark, the model scores 92.7 percent with tools. This compares to 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra. Without tools, it reaches 72.8 percent, against 73.0 percent for Claude Opus 5.5. A new 3D Make feature on Kimi.com allows users to preview, edit, and export 3D models.
The architecture remains unchanged from K3, featuring 2.8 trillion total parameters, 104 billion of which are active. The model supports a one-million-token context window and handles both text and images. Artificial Analysis places it at the top of the Pareto frontier, indicating that no model of comparable size scores higher on its index. API prices have fallen by 33 percent. Input costs stand at $2 per million tokens, output at $10, and cache hits at $0.20. Although the model generates 14 percent more output tokens, the cost of running each Intelligence Index task is $1.98, nearly identical to K3's $2.00.
Moonshot addressed cybersecurity in its technical blog post. The company states that vulnerability research capabilities grew as post-training scaled up. This occurred because cyber-related content was part of the general mid-training corpus and related tasks were included in the reinforcement learning mix. For the open-weights release, the company removed cyber-related data from both training phases. On its internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, and JavaScriptCore, the uncleaned model solved 90.5 percent of tasks. This is versus 81.0 percent after cleaning. DeepSeek-V4.1-Flash scored 85.7 percent, while Kimi K3 scored 33.3 percent. Moonshot argues that open-weights models should not push the absolute frontier of offensive cyber capabilities, but it aims to raise their potential gradually to support defensive work like auditing legacy systems. Anthropic previously demonstrated that even GLM-5.3, which scores 45 on the index, can write exploits at levels close to Claude Mythos.
Source: The Decoder
