New Chinese AI Model Narrows Gap With Top US Systems
Moonshot AI's latest release has cut the performance distance between open-source and proprietary models to just three points.
By Kateryna Melnyk ·
A three-point gap now separates the world's most capable artificial intelligence model from an offering released by a Chinese developer this week, a margin that underscores how quickly the competitive landscape is shifting. Moonshot AI's Kimi K3.1 Pro reached a score of 55 on the Artificial Analysis Intelligence Index, a jump of 11 points from the company's prior model. The release places the system at number four on the global rankings. The company announced the model and its pricing on Wednesday.
The scoring system places K3.1 Pro directly behind Claude Opus 5.5, which sits at 58. Two other systems, Claude Sonnet 5.5 and Muse Wonder Full, are both at 56. Moonshot plans to release the model weights by November 4, 2026. If that schedule holds, K3.1 Pro would become the highest-scoring open-weights model available on the market. The gap between open and proprietary systems would shrink from 12 points to three, assuming no other company releases a higher-scoring model before that date. The model is available now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, along with a variant called K3.1 Pro Highspeed that runs two to three times faster.
The model shows its largest gains in agentic tasks, which involve a system working through steps to complete a request. On the GDPval-AA benchmark, the model reached an Elo of 1,846, a rise of 312 points from the prior version. That result trails Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On the Terminal-Bench 4.0 test, the system scored 65.7 percent, ahead of GPT-6 Astra at 58.2 percent, and just behind Claude Opus 5.5 at 66.4 percent. The prior version of the model managed only 13 percent on that same test. Claim-level hallucination rates fell from 23.4 percent to 8.3 percent on an internal test, though GPT-6 Astra hits 5.5 percent on the same measure.
The company also reports progress in coding and 3D modeling. On the BenchCAD benchmark, the model scores 92.7 percent with tools, close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Without tools, it reaches 72.8 percent, nearly matching Claude Opus 5.5 at 73.0 percent. The model can write and run scripts for Blender and CAD software without connectors, and it works directly with game engines such as Unity, Unreal Engine, and Godot. On internal financial benchmarks, the model gained 7.2 points on Financial Modeling and 11.6 points on Financial Research. For general office tasks, Moonshot reports the model tops all tested systems on OfficeQA Pro v2.
The model holds the same size as its predecessor, with 2.8 trillion total parameters and 104 billion active ones, while token prices drop by 33 percent. Input costs $2 per million tokens, down from $3. Output costs $10, down from $15, and cache hits cost $0.20, down from $0.30. The cost per task on the Intelligence Index is $1.98, nearly identical to the prior model at $2.00, since the system generates 14 percent more output tokens.
Moonshot also addressed cybersecurity in its technical documentation. The company removed cyber-related data from both training phases for the open-weights release. On an internal test that reproduces recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, and JavaScriptCore, the version without data cleaning solves 90.5 percent of tasks. After data cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on that test, while Kimi K3 scored 33.3 percent. Moonshot states that no open-weights model should advance the absolute frontier of offensive cyber capabilities, though the company wants to raise what open models can do for defensive purposes such as auditing legacy systems.
The release arrives as the cost of competing in this market continues to shift. Lower token prices and the planned release of model weights would put a high-scoring system within reach of developers who cannot pay for proprietary access. The narrow margin between this model and the top-ranked systems suggests the gap in capability between open and closed systems is closing, a development that could affect how companies choose to build products on these platforms.
Source: The Decoder
