Moonshot AI Releases Kimi K3.1 Pro With Lower Prices

Moonshot AI has launched Kimi K3.1 Pro. This is a new reasoning model that aims to close the gap with leading American competitors. The release focuses on performance and price. It is a post-trained version of the previous Kimi K3. The company used significantly more compute for this training run. The goal is to improve specific technical tasks. This includes coding and complex agent work. The model is available to the public now. It is accessible through several of Moonshot's platforms. This includes the Kimi API and Kimi Code. A faster variant called K3.1 Pro Highspeed is also out. It runs two to three times faster than the standard version.

The new model scored 55 points on the Artificial Analysis Intelligence Index. This marks an 11-point jump over the previous version. It now holds the fourth position on the leaderboard. Only three models rank higher right now. Those are Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. The scores for those three are 58, 56, and 56 respectively. The gap between Kimi K3.1 Pro and the top models is just three points. Moonshot plans to release the model weights by November 4, 2026. If that happens, it will be the highest-scoring open-weights model. This would shrink the difference between open and proprietary models. The gap would drop from 12 points to just 3. This assumes no other major releases happen before then.

The biggest improvements appear in agentic tasks. On the GDPval-AA benchmark, the model reached 1,846 Elo. That is 312 points higher than the previous K3. It sits just behind Claude Opus 5.5 and Muse Wonder Full. On Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent. This beats GPT-6 Astra, which scored 58.2 percent. It is also very close to Claude Opus 5.5 at 66.4 percent. The older K3 model managed only 13 percent on that same test. Hallucination rates have also dropped. On an internal stress test, claim-level errors fell from 23.4 percent to 8.3 percent. GPT-6 Astra still has an edge with a 5.5 percent rate. However, Moonshot notes this is a hard test, not a general measure.

The model shows strong gains in finance and office work. It gained 7.2 points on financial modeling tasks. It also gained 11.6 points on financial research. For general office tasks, it tops all tested models on OfficeQA Pro v2. The company says these results show better handling of professional duties. The model now works directly with game engines. It can handle Unity, Unreal Engine, and Godot. It manages both quick fixes and full project builds. This is a significant step for developers using these tools. It reduces the need for manual setup in many cases. The integration is designed to be seamless for users.

Coding and 3D modeling capabilities have improved notably. On the BenchCAD benchmark, the model scores 92.7 percent with tools. This is close to Claude Opus 5.5 at 96.2 percent. GPT-6 Astra scores 95.9 percent on the same test. Without tools, Kimi K3.1 Pro hits 72.8 percent. Claude Opus 5.5 scores 73.0 percent in that scenario. Moonshot says the model excels at parametric modeling. It handles mechanical parts and architectural models well. It can write and run scripts for Blender and CAD software. It does this without needing MCP connectors. A new feature called 3D Make is also live on Kimi.com. Users can preview, edit, and export 3D models there.

The model size stays the same as before. It has 2.8 trillion total parameters. Of those, 104 billion are active at any time. The context window spans one million tokens. It handles both text and images. Artificial Analysis places it at the top of the Pareto frontier. No model with a similar parameter count scores higher. This efficiency is a key selling point. The company kept the architecture stable. This allows for predictable performance scaling. It also simplifies deployment for businesses. The consistent size makes upgrades easier for existing users. It reduces the need for major infrastructure changes.

Prices have dropped significantly for the new model. Token costs are down by 33 percent. Input costs are now $2 per million tokens. This was $3 before. Output costs are $10 per million tokens. That was $15 previously. Cache hits cost $0.20 per million tokens. That was $0.30 earlier. The cost per Intelligence Index task is $1.98. This is almost identical to the previous cost of $2.00. This is because the model generates 14 percent more output tokens. The lower price point makes it attractive for high-volume use. It offers better value for money. This could pressure competitors to adjust their pricing strategies.

Moonshot addressed cybersecurity in its technical report. It noted that vulnerability research capabilities grew during training. This happened because cyber content was in the training data. For the open-weights release, the company removed this data. It cleaned both training phases to be safe. On an internal test for recent vulnerabilities, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash scores 85.7 percent on that test. The older Kimi K3 scored only 33.3 percent. Moonshot wants to raise what open models can do. It believes they serve defensive purposes too. But it does not want them to push the absolute frontier of offensive cyber capabilities.

The stakes are high for the AI market. Open models are getting closer to proprietary ones. The gap is narrowing fast. This could change how companies choose their AI tools. Price and performance are the main drivers. Moonshot is making a strong move here. It offers top-tier performance at a lower cost. The November release of weights is a key date to watch. If it holds, open models will lead in score. This shifts the balance of power in the industry. Developers will have more options. They can choose between closed and open systems. The competition is heating up. Users benefit from this rivalry. Better tools and lower prices are the result.