Kimi K3.1 Pro Narrows Gap With Top Rivals

Moonshot AI has released Kimi K3.1 Pro, a new reasoning model that has jumped to fourth place on the Artificial Analysis Intelligence Index. The model scores 55 points, placing it just behind the leading proprietary systems. This marks an 11 point increase over its predecessor, Kimi K3. The gap between this new Chinese release and the top tier is now very small. Only three other models rank higher according to current data. These are Claude Opus 5.5 with 58 points, Claude Sonnet 5.5 with 56 points, and Muse Wonder Full also at 56 points.

If Moonshot follows through on its plan to release the model weights by November 4, 2026, the dynamics of the market will shift. Kimi K3.1 Pro would become the highest scoring open weights model available. That move would shrink the performance gap between open and proprietary models from 12 points to just three. This assumes that no other competing models improve their scores before that date. The release would represent a significant milestone for open source AI capabilities.

The company is also shipping a variant called K3.1 Pro Highspeed. According to Moonshot, this version runs two to three times faster than the standard model. Both versions are available immediately through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Users in New England and elsewhere can access these tools now without waiting for a wider distribution rollout.

The most dramatic improvements are seen in agentic tasks. On the GDPval-AA benchmark, the new model reaches an Elo score of 1,846. This is a 312 point jump over the previous K3 model. It places the system 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, and is close to Claude Opus 5.5 at 66.4 percent. The previous K3 model managed only 13 percent on this specific test.

Hallucination rates have dropped significantly as well. On Moonshot's internal Kimi Hallu Bench Hard, claim level hallucinations fell from 23.4 percent in K3 to 8.3 percent. The company notes that this is a stress test rather than a measure of general error rates. GPT-6 Astra still holds an edge with a score of 5.5 percent on the same benchmark. In internal finance benchmarks, K3.1 Pro gained 7.2 points in Financial Modeling and 11.6 points in Financial Research. The model also tops all tested rivals on OfficeQA Pro v2 for general office tasks.

Capabilities in 3D modeling and coding have expanded. The model works directly with game engines like Unity, Unreal Engine, and Godot. It can handle quick fixes as well as full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools. This is close to Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. Without tools, the score is 72.8 percent, which nearly matches Claude Opus 5.5 at 73.0 percent. Moonshot says the model excels at parametric modeling of mechanical parts and architectural models. It can write and run scripts for Blender and CAD software without needing MCP connectors. A new feature called 3D Make on Kimi.com allows users to preview, edit, and export 3D models.

The model keeps the same dimensions as K3, with 2.8 trillion total parameters and 104 billion active ones. The context window spans one million tokens, and it handles both text and images. Artificial Analysis places it at the top of the Pareto frontier for its size. Token prices have dropped by 33 percent overall. Input costs are now $2 per million tokens, down from $3. Output costs are $10, down from $15. Cache hits cost $0.20, down from $0.30. Each Intelligence Index task costs $1.98, which is almost identical to the previous model at $2.00. This is because the new model generates 14 percent more output tokens.

Moonshot addressed cybersecurity in its technical blog post. The company observed that vulnerability research capabilities grew as post training scaled up. This happened because cyber related content was in the mid training corpus and RL environment. For the open weights release, they stripped cyber related data from both training phases. On an internal test called xdev-bench-hard, the uncleaned version solved 90.5 percent of tasks. After cleaning, K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash reaches 85.7 percent on the same test. Moonshot believes open weights models should not push the absolute frontier of offensive cyber capabilities but should gradually raise what is possible for defensive purposes.