The release places Moonshot among the highest-ranked AI model developers in independent evaluations. Artificial Analysis awarded Kimi K3.1 Pro a score of 55 on its Intelligence Index, an 11-point improvement over K3 that puts the new model in fourth place overall.
Anthropic's Claude Opus 5.5 leads the comparison with 58 points, followed by Claude Sonnet 5.5 and Muse Wonder Full at 56 each. Artificial Analysis expects Kimi K3.1 Pro to become the highest-scoring open-weight model once its weights are publicly available, reducing the performance difference between the leading open-weight and proprietary models from 12 points to three.
Much of the improvement comes from tasks requiring AI agents to work through multiple steps and interact with development tools. Artificial Analysis recorded a 67% result for Kimi K3.1 Pro on Terminal-Bench 4.0, compared with 13% for its predecessor.
Moonshot's own testing produced a slightly different result of 65.7%, placing the model ahead of OpenAI's GPT-6 Astra at 58.2% and close to Claude Opus 5.5 at 66.4%. These company-reported comparisons have not been independently verified.
On GDPval-AA, an evaluation of professional work tasks, Kimi K3.1 Pro achieved an Elo rating of 1,846, placing third behind Claude Opus 5.5 at 1,866, according to Artificial Analysis.
Moonshot is positioning the model for software development, 3D design and financial analysis. The company says K3.1 Pro can complete larger coding projects, including game development using Unity, Unreal Engine and Godot. Its new 3D Make feature, available through the Kimi website and mobile app, allows users to generate, preview and modify three-dimensional models, including designs that can be downloaded for 3D printing.
The company also reported improvements across specialized technical evaluations. On BenchCAD, which measures the ability to reconstruct 3D objects using generated CAD code, K3.1 Pro scored 92.7% with tools. Moonshot reported a score of 64.4 on OfficeQA Pro v2, which evaluates analysis of complex financial documents, while its internal Business Operation benchmark rose from 8.9 for K3 to 37 for the new model.
Accuracy was another focus of the upgrade. Moonshot's Kimi Hallu Bench Hard evaluation found that the proportion of incorrect factual claims declined from 23.4% with K3 to 8.3% with K3.1 Pro when tools were available. The company describes this benchmark as a demanding test of situations prone to factual errors rather than a measure of typical everyday performance.
Independent testing by Artificial Analysis also showed improvement, although the results were less pronounced. Its AA-Omniscience evaluation recorded a reduction in hallucination rate from 53% to 44%.
Alongside the performance gains, Moonshot disclosed an unusual approach to limiting the cybersecurity capabilities of its open-weight release. The company removed material associated with cyberattacks from portions of the model's mid-training data and reinforcement-learning exercises while retaining general software development training.
Moonshot said the changes were intended to prevent the publicly downloadable model from establishing a new performance frontier in offensive cybersecurity.
In its internal xdev-bench-hard evaluation, which measures the ability to identify and exploit software vulnerabilities, the final model completed 81% of tasks. A comparable checkpoint trained without the cybersecurity data restrictions achieved 90.5%, while the earlier K3 model scored 33.3%.
Moonshot also reported a score of 85.7% for DeepSeek V4.1-Flash on the same evaluation. The company maintains that K3.1 Pro's offensive cybersecurity capabilities remain below those of leading proprietary systems, although its internal benchmark results have not been independently confirmed.
The distinction is particularly relevant to an open-weight release because downloaded model weights can be used outside the developer's hosted services. Moonshot says its approach is intended to balance wider access to cybersecurity capabilities with limits on the most advanced offensive techniques.
For developers, Kimi K3.1 Pro is accessible through the API under the model identifier kimi-k3.1-pro, while the faster option is available as kimi-k3.1-pro-highspeed. Both offer Zero Data Retention support.
API pricing starts at $2 per million input tokens and $10 per million output tokens, with cached input priced at $0.20 per million tokens. These rates are 33% below those of Kimi K3, according to Artificial Analysis.
However, the lower token prices do not translate directly into equivalent savings for every workload. Artificial Analysis found that K3.1 Pro generates approximately 14% more output tokens, bringing its average cost per Intelligence Index task to $1.98, roughly comparable to its predecessor.
The models are also available through Kimi's web and mobile applications, Kimi Work and Kimi Code. Enterprise administrators can activate K3.1 Pro for their organizations, although access is disabled by default at launch. In Kimi Code, the available context length and model selection depend on membership tier.
Moonshot's release combines substantial improvements in agentic performance with lower API pricing and a planned open-weight distribution. The independent benchmark results put Kimi K3.1 Pro close to several leading proprietary models, while its cybersecurity training changes reflect an effort to limit certain capabilities before making the weights available.
The next milestone is the planned November 4 publication of the complete model weights, which would allow developers to run and examine Kimi K3.1 Pro beyond Moonshot's hosted services.
This analysis is based on reporting from TNW.
Image courtesy of kimi.ai.
This article was generated with AI assistance and reviewed for accuracy and quality.
