Moonshot AI cuts prices, boosts performance

Moonshot AI has launched Kimi K3.1 Pro, a post-trained reasoning model that significantly narrows the competitive distance to Anthropic’s leading products. The new release scores 55 points on the Artificial Analysis Intelligence Index, placing it in fourth position globally. This result trails only Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full, marking an 11-point improvement over its predecessor, Kimi K3. The company positions the update as a major step toward parity between Chinese and American frontier models, particularly in complex reasoning and agentic workflows.

The competitive landscape shifted rapidly in the weeks leading up to this launch. Moonshot previously released Kimi K3, which established a baseline for the company’s large language models. However, the gap between open-weights and proprietary systems remained significant, with a 12-point difference in intelligence scores. By increasing post-training compute, Moonshot has addressed this disparity. If the company executes its plan to release model weights by November 4, 2026, K3.1 Pro will become the highest-scoring open-weights model available. This move would shrink the gap between open and proprietary leaders to just three points, assuming no other major releases occur before then.

Performance gains are most evident in agentic tasks and coding. On the GDPval-AA benchmark, the model reaches 1,846 Elo, a 312-point jump from K3. It now sits just behind Claude Opus 5.5 and Muse Wonder Full. In terminal-based coding tests, specifically Terminal-Bench 4.0, K3.1 Pro achieves a 65.7 percent success rate. This surpasses GPT-6 Astra, which scored 58.2 percent, and nearly matches Claude Opus 5.5 at 66.4 percent. Notably, the previous K3 model managed only 13 percent on the same test, indicating a substantial leap in reliability for software engineering tasks.

The company also reports a sharp reduction in hallucinations. On its internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent with K3 to 8.3 percent. While Moonshot notes this is a stress test rather than a measure of general error rates, the improvement is significant. GPT-6 Astra currently holds an edge with a 5.5 percent rate on the same benchmark. Additionally, K3.1 Pro shows strong results in finance, gaining 7.2 points on Financial Modeling and 11.6 points on Financial Research. It also tops all tested models on OfficeQA Pro v2 for general office tasks.

A distinctive feature of the new release is its capability in 3D modeling. K3.1 Pro works directly with game engines such as Unity, Unreal Engine, and Godot, handling both minor fixes and full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools, closely following Claude Opus 5.5 and GPT-6 Astra. Without tools, it hits 72.8 percent, nearly matching Claude Opus 5.5. Moonshot states the model excels at parametric modeling of mechanical parts and architectural structures. It can also write and run scripts for Blender and CAD software without requiring MCP connectors, and the company has launched a "3D Make" feature on Kimi.com to allow users to preview and export models.

From a market perspective, Moonshot has maintained the model’s architectural size while reducing costs. K3.1 Pro retains 2.8 trillion total parameters and 104 billion active ones, with a context window of one million tokens. However, API prices have dropped by 33 percent. Input costs are now $2 per million tokens, down from $3, while output costs fall to $10 from $15. Cache hits cost $0.20, reduced from $0.30. According to Artificial Analysis, this places the model at the top of the Pareto frontier, as no other model with a comparable parameter count scores higher on the Intelligence Index. The cost per Intelligence Index task is $1.98, nearly identical to K3, due to a 14 percent increase in output tokens.

The company has also adjusted its approach to cybersecurity capabilities. Moonshot observed that vulnerability research skills grew as post-training scaled, due to cyber-related content in the training corpus. For the upcoming open-weights release, the company stripped this data from both training phases. On its internal xdev-bench-hard test, which reproduces recently disclosed vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent, compared to 85.7 percent for DeepSeek-V4.1-Flash. Moonshot aims to balance offensive potential with defensive utility, noting that open models also serve purposes like auditing legacy systems.

Looking ahead, both K3.1 Pro and the faster K3.1 Pro Highspeed variant are available immediately through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The Highspeed variant runs two to three times faster, according to the company. The key next milestone is the scheduled release of model weights in November 2026. This will allow developers to fine-tune the model locally, potentially altering the competitive dynamics in the AI market. For now, Moonshot has provided a cost-effective option that challenges the dominance of U.S.-based providers in specific technical domains.