Moonshot AI Releases Kimi K3.1 Pro, Narrowing the Gap with Top US Models

Moonshot AI has released Kimi K3.1 Pro, a post-trained reasoning model that has surged to the fourth position on the Artificial Analysis Intelligence Index with a score of 55. This 11-point improvement over its predecessor, Kimi K3, places the new model just three points behind the leading proprietary systems, including Anthropic’s Claude Opus 5.5, which scores 58. The release signals a significant acceleration in the capabilities of Chinese AI development, narrowing the performance divide between open and closed models to a single-digit margin.

The primary value of this release lies in its efficiency and cost structure. While maintaining the same architectural scale of 2.8 trillion total parameters and 104 billion active ones, Moonshot has reduced API pricing by 33 percent. Input costs now stand at $2 per million tokens, down from $3, while output costs have dropped to $10 from $15. This price reduction, combined with the model’s high ranking, positions K3.1 Pro at the top of the Pareto frontier, meaning no other model with a comparable parameter count achieves a higher intelligence score. The company also introduced a 'Highspeed' variant that operates two to three times faster, both of which are currently available through various Moonshot platforms.

Performance gains are most evident in agentic tasks and coding. On the GDPval-AA benchmark, K3.1 Pro achieved an Elo rating of 1,846, a 312-point increase over K3. In Terminal-Bench 4.0, the model recorded a 65.7 percent success rate, surpassing GPT-6 Astra and coming within less than one percent of Claude Opus 5.5. Additionally, the model demonstrates enhanced capabilities in 3D modeling, scoring 92.7 percent on BenchCAD with tools, and can directly interact with game engines like Unity and Unreal Engine. Moonshot has also launched a '3D Make' feature on its website, allowing users to preview and export generated models.

The company has also emphasized improvements in reliability and security. Internal testing shows that claim-level hallucination rates dropped from 23.4 percent to 8.3 percent, though this remains slightly higher than GPT-6 Astra’s 5.5 percent. Regarding cybersecurity, Moonshot noted that vulnerability detection capabilities grew naturally with post-training. To balance offensive and defensive potential for the open-weights release, the company removed specific cyber-related data from its training corpus. This adjustment reduced the model’s score on internal vulnerability benchmarks from 90.5 percent to 81.0 percent, a deliberate choice to prevent open models from dominating offensive capabilities while still aiding in defensive auditing.

If Moonshot adheres to its plan to release the model weights by November 4, 2026, Kimi K3.1 Pro will become the highest-scoring open-weights model available. This would shrink the gap between the best open and proprietary models from 12 points to just three, assuming no other major releases occur in the interim. The move underscores a strategic shift in the AI landscape, where cost-efficiency and agentic reliability are becoming as critical as raw benchmark scores.

The release includes a one-million-token context window and support for both text and image inputs. While the model generates 14 percent more output tokens than its predecessor, the overall cost per Intelligence Index task remains nearly identical at $1.98. As the open-weights version approaches, the competitive pressure on established US firms like Anthropic and OpenAI is set to intensify, particularly in enterprise applications that require robust coding and 3D generation capabilities.