55 Points: Moonshot AI's Kimi K3.1 Pro Narrows Gap With Top Proprietary Models

Moonshot AI has released Kimi K3.1 Pro, a model that has surged to fourth place on the Artificial Analysis Intelligence Index with a score of 55. This 11-point increase over its predecessor, Kimi K3, positions it just three points behind the current leaders, including Anthropic’s Claude Opus 5.5. The release marks a significant closing of the gap between high-performing proprietary systems and emerging competitors in the global AI landscape.

The new model delivers substantial improvements in agentic tasks, coding, and 3D modeling. On the GDPval-AA benchmark, K3.1 Pro achieves an Elo rating of 1,846, a 312-point jump from its previous version. It also records a 65.7 percent score on Terminal-Bench 4.0, a massive improvement from the 13 percent recorded by K3. These figures place it nearly on par with Claude Opus 5.5 and ahead of GPT-6 Astra in specific terminal-based evaluations.

Despite the performance gains, the model retains its 2.8 trillion total parameters and 104 billion active parameters. Moonshot has reduced API pricing by 33 percent, lowering input costs to $2 per million tokens and output costs to $10. This strategic move keeps the model at the top of the Pareto frontier, meaning no other model with a similar parameter count scores higher on the Intelligence Index. The company plans to release the model weights by November 4, 2026, which would make it the highest-scoring open-weights model available.

K3.1 Pro also demonstrates enhanced capabilities in 3D modeling and finance. It can work directly with game engines like Unity and Unreal Engine, scoring 92.7 percent on the BenchCAD benchmark with tools. In financial tasks, the model gained 11.6 points on research benchmarks. Furthermore, hallucination rates have dropped significantly, with claim-level errors falling from 23.4 percent to 8.3 percent on internal stress tests, although GPT-6 Astra still maintains a lower error rate.

For the upcoming open-weights release, Moonshot has deliberately limited offensive cybersecurity capabilities to balance innovation with safety. By stripping cyber-related data from the training corpus, the model’s ability to solve vulnerability tasks dropped from 90.5 percent to 81.0 percent. This approach aims to prevent open models from pushing the absolute frontier of offensive cyber capabilities while still serving defensive purposes like auditing legacy systems. A high-speed variant, K3.1 Pro Highspeed, is also available now, running two to three times faster than the standard version.