Moonshot AI Launches K3.1 Pro to Challenge Top AI Models

Moonshot AI has released Kimi K3.1 Pro, a new reasoning model that now ranks fourth on the Artificial Analysis Intelligence Index with a score of 55. This result marks an 11 point improvement over its predecessor, Kimi K3, and places the system just three points behind the highest scoring proprietary models currently available.

K3.1 Pro is a post trained version of K3 that utilized substantially more post training compute. According to data from Artificial Analysis, only three models currently outperform it. These include Anthropic's Claude Opus 5.5 with a score of 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56. Moonshot intends to publish the model weights by November 4, 2026. If this timeline holds, K3.1 Pro would become the highest scoring open weights model. This move would reduce the performance gap between open and proprietary systems from 12 points to just three, assuming no other changes occur in the interim.

Alongside the main release, the company is offering a variant called K3.1 Pro Highspeed. Moonshot states that this version operates two to three times faster than the standard model. Both versions are accessible through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API platform.

The most significant gains occurred in agentic tasks. On the GDPval AA benchmark, K3.1 Pro achieved an Elo rating of 1,846. This is 312 points higher than the previous K3 model and sits slightly behind Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. In terminal based evaluations, the model recorded a 65.7 percent score on Terminal Bench 4.0. This places it ahead of GPT 6 Astra at 58.2 percent and marginally below Claude Opus 5.5 at 66.4 percent. For context, the older K3 model managed only 13 percent on that specific test.

The company also reported 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. Moonshot describes this benchmark as a stress test rather than a measure of general error rates. It notes that GPT 6 Astra still leads this metric with a rate of 5.5 percent. Internal finance benchmarks showed similar progress, with the model gaining 7.2 points on Financial Modeling and 11.6 points on Financial Research. Moonshot claims K3.1 Pro currently tops all tested models on OfficeQA Pro v2.

Capabilities in 3D modeling and coding were also highlighted. The model interacts with game engines such as Unity, Unreal Engine, and Godot. It can write and run scripts for Blender and CAD software without requiring MCP connectors. On the BenchCAD benchmark, K3.1 Pro scored 92.7 percent with tools, compared to 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT 6 Astra. Without tools, it reached 72.8 percent, a figure very close to the 73.0 percent achieved by Claude Opus 5.5. A new 3D Make feature has been added to Kimi.com to allow users to preview, edit, and export models.

The underlying architecture remains unchanged from K3. It features 2.8 trillion total parameters, with 104 billion active, a one million token context window, and support for text and images. Artificial Analysis places the model on the top of the Pareto frontier, indicating that no model of comparable size scores higher on its index. API prices have decreased by 33 percent. Input costs are now $2 per million tokens, output is $10, and cache hits cost $0.20. Because the model generates 14 percent more output tokens, the cost of running each Intelligence Index task is $1.98, which is nearly identical to the $2.00 cost of K3.

Moonshot dedicated a portion of its technical documentation to cybersecurity. The company noted that vulnerability research capabilities improved as post training scaled up. This was because cyber related content was part of the general mid training corpus and related tasks were included in the reinforcement learning mix. For the open weights release, cyber related data was removed from both training phases. On an internal benchmark that reproduces recently disclosed vulnerabilities in software like the Linux kernel and Node.js, the uncleaned model solved 90.5 percent of tasks. After cleaning, this figure dropped to 81.0 percent. DeepSeek V4.1 Flash scored 85.7 percent, while Kimi K3 achieved 33.3 percent.

The company stated that no open weights model should push the absolute frontier of offensive cyber capabilities. However, it aims to raise what open models can do gradually to support defensive work such as auditing legacy systems. Anthropic previously demonstrated that even GLM 5.3, which scores 45 on the Artificial Analysis index, can write exploits at levels close to Claude Mythos.