Moonshot's New AI Model Scores 55 Points, Nearing Top Tier
A new reasoning model from a Chinese lab has leaped ahead in performance while cutting costs significantly, challenging the dominance of established Western competitors.
By Juan Gonzalez ·
Moonshot AI has just closed the gap with the industry's best. The company released Kimi K3.1 Pro, a new reasoning model that immediately jumped to fourth place on the Artificial Analysis Intelligence Index with a score of 55 points. This places it just three points behind the top proprietary models.
The performance metrics show substantial improvements across the board. K3.1 Pro recorded an 11-point jump over its predecessor, Kimi K3. On the GDPval-AA benchmark, the new model reached 1,846 Elo, surpassing the previous version by 312 points. It also achieved a 65.7 percent score on Terminal-Bench 4.0, a massive increase from the 13 percent recorded by K3. Only Claude Opus 5.5 and Muse Wonder Full currently outperform it in these specific agentic tasks.
The model is immediately available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Moonshot also introduced a high-speed variant that runs two to three times faster. The core architecture retains the 2.8 trillion parameter count of the previous version, with 104 billion active parameters and a one-million token context window. Despite keeping the same size, the company cut API prices by 33 percent. Input costs are now $2 per million tokens, while output costs sit at $10.
The model demonstrates particular strength in 3D modeling and coding. It can interact directly with game engines like Unity and Unreal Engine. On the BenchCAD benchmark, it scored 92.7 percent when using tools, trailing only Claude Opus 5.5 and GPT-6 Astra. Moonshot claims it is especially effective at parametric modeling for mechanical and architectural parts. A new feature called 3D Make on Kimi.com allows users to preview and export these models.
Hallucination rates have also dropped significantly. According to the company's internal stress tests, claim-level hallucinations fell from 23.4 percent in K3 to 8.3 percent in the new release. While GPT-6 Astra still holds an edge with a 5.5 percent rate, the reduction indicates a tighter focus on accuracy. Moonshot states that the model tops all tested competitors on the OfficeQA Pro v2 benchmark for general office tasks.
The stakes are high for the open-weights AI sector. If Moonshot releases the model weights by November 4, 2026, as planned, K3.1 Pro will become the highest-scoring open model available. This would shrink the performance gap between open and proprietary models from 12 points to just three. The company noted that while it stripped some cyber-related data from the training set for the open release, the model still solves 81.0 percent of tasks on its internal security benchmark. This move suggests a deliberate strategy to balance defensive utility with offensive capability.
This release intensifies competition at the top of the AI ladder. With lower costs and near-frontier performance, K3.1 Pro challenges the pricing and performance structures established by major Western labs. The shift highlights a rapid evolution in model capabilities, where significant gains in agentic tasks and coding are now being paired with aggressive pricing strategies and open-weight plans.
Source: The Decoder
