Moonshot AI releases Kimi K3.1 Pro to challenge top US models

For businesses and developers in Texas and the South Central states relying on artificial intelligence, a new option from China just got significantly cheaper and more capable. Moonshot AI has released Kimi K3.1 Pro, a model that now sits just behind the leading proprietary systems from the United States. This update means teams can access high-level reasoning and coding power at a lower cost per token than before.

The model currently holds the fourth spot on the Artificial Analysis Intelligence Index with a score of 55. It is only three points behind Claude Opus 5.5, which leads the pack with 58 points. Also ranking higher are Claude Sonnet 5.5 and Muse Wonder Full, both at 56 points. This places Kimi K3.1 Pro firmly in the top tier of global AI performance.

Moonshot plans to release the model weights by November 4, 2026. If that timeline holds, K3.1 Pro would become the highest-scoring open-weights model available. This would narrow the gap between open and proprietary models from 12 points to just 3. Alongside the main release, the company is offering a variant called K3.1 Pro Highspeed. This version runs two to three times faster and is available now through Kimi.ai, Kimi Work, Kimi Code, and the API.

The biggest improvements in K3.1 Pro appear in agentic tasks, which involve the model performing multi-step actions. On the GDPval-AA benchmark, the model reaches an Elo score of 1,846. That is a 312-point increase over its predecessor, Kimi K3. It now trails only Claude Opus 5.5 and Muse Wonder Full on that specific test. In terminal-based coding tasks, K3.1 Pro scores 65.7 percent on Terminal-Bench 4.0. This is a massive jump from the 13 percent recorded by K3 and beats GPT-6 Astra, which scored 58.2 percent.

The model also shows better accuracy in its factual claims. On an internal stress test called Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent to 8.3 percent. Moonshot notes this is a rigorous test rather than a measure of general error rates. GPT-6 Astra still performs better on this specific benchmark with a 5.5 percent hallucination rate. In financial applications, K3.1 Pro gained 7.2 points on Financial Modeling and 11.6 points on Financial Research compared to previous versions.

A notable strength for this release is its ability to handle 3D modeling and game development. The model works directly with engines like Unity, Unreal Engine, and Godot. It can handle everything from quick fixes to full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using external tools. Without tools, it achieves 72.8 percent, which is nearly on par with Claude Opus 5.5. Moonshot says the model excels at parametric modeling for mechanical parts and architecture.

Despite the performance gains, the underlying size of the model remains unchanged. It keeps 2.8 trillion total parameters with 104 billion active ones. The context window spans one million tokens and supports both text and images. According to Artificial Analysis, it sits at the top of the Pareto frontier for its size class. No model with a comparable parameter count scores higher on the Intelligence Index.

The most immediate benefit for users is the price cut. Token prices have dropped by 33 percent across the board. Input costs are now $2 per million tokens, down from $3. Output costs are $10 per million tokens, reduced from $15. Cache hits cost $0.20 per million tokens, down from $0.30. Because the model generates 14 percent more output tokens, the cost per task remains almost identical to the previous version at $1.98.

Moonshot also addressed cybersecurity in its technical documentation. The company found that vulnerability research capabilities grew as post-training scaled up. For the open-weights release, they removed cyber-related data from both training phases to manage offensive capabilities. On an internal test reproducing recent software vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After cleaning the data, K3.1 Pro solves 81.0 percent. This is higher than Kimi K3, which solved 33.3 percent, but lower than DeepSeek-V4.1-Flash, which reached 85.7 percent. The company aims to gradually raise what open models can do for defensive purposes while keeping offensive capabilities in check.