Moonshot AI: New model narrows the gap with top global competitors

Moonshot AI has just rolled out Kimi K3.1 Pro, a new reasoning model that lands in fourth place on the Artificial Analysis Intelligence Index. It scores 55 points, sitting just behind three leading proprietary systems. The company describes this as a post-trained version of its previous Kimi K3 model, which utilized significantly more computational power during the final training stages.

The performance gains are substantial, particularly in agentic tasks. On the GDPval-AA benchmark, the new model reaches an Elo rating of 1,846. This is a jump of 312 points over its predecessor. It trails only Claude Opus 5.5 and Muse Wonder Full on this specific metric. Coding abilities have also improved dramatically. On Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent. This far exceeds the 13 percent score achieved by the older K3 model.

Reliability has seen a boost as well. Internal tests show a sharp decline in hallucinations. On Moonshot's Kimi Hallu Bench Hard, claim-level errors dropped from 23.4 percent to 8.3 percent. The company notes this is a stress test rather than a general measure of error rates. GPT-6 Astra still holds an edge with a 5.5 percent rate, but the improvement for Kimi is notable for users seeking consistent output.

For creative and technical users, the model now handles 3D modeling and game engine workflows. It works directly with Unity, Unreal Engine, and Godot. On the BenchCAD benchmark, it scores 92.7 percent when using tools. Without external tools, it hits 72.8 percent, which nearly matches the performance of Claude Opus 5.5. A new feature called 3D Make allows users to preview and export models directly on the Kimi.com platform.

Pricing and accessibility have also improved. The model retains 2.8 trillion total parameters but offers API prices that are 33 percent lower than before. Input costs are now $2 per million tokens, down from $3. Output costs stand at $10 per million, reduced from $15. Moonshot plans to release the model weights by November 4, 2026. If nothing changes, this would make it the highest-scoring open-weights model available.

Security considerations played a role in the development process. Moonshot stripped cyber-related data from the training phases for the open-weights release. This decision was made to balance offensive and defensive capabilities. The company believes open models should not push the absolute frontier of offensive cyber capabilities but should still serve defensive purposes like auditing legacy systems. The uncleaned version solved 90.5 percent of tasks on their internal security bench, while the cleaned version solved 81.0 percent.

Both the standard and high-speed variants of the model are available now through various Moonshot platforms. The high-speed version runs two to three times faster according to the company. This release signals a continued push to close the performance gap between open and proprietary AI systems while keeping costs manageable for developers and enterprises.