New Chinese AI model cuts price and boosts performance

A 33 percent drop in API pricing is the headline figure accompanying the release of Kimi K3.1 Pro, a new reasoning model from Chinese AI firm Moonshot AI. The company announced the update on October 23, 2026, highlighting that the model now costs two dollars per million input tokens, down from three dollars, while output costs have fallen from fifteen dollars to ten dollars per million tokens. Cache hit prices have also decreased to twenty cents per million tokens from thirty cents.

The performance improvements accompanying the price cut are substantial. Kimi K3.1 Pro has climbed to fourth place on the Artificial Analysis Intelligence Index with a score of 55. This represents an 11-point increase over its predecessor, Kimi K3. It now trails only Claude Opus 5.5, which holds the top spot with 58 points, followed by Claude Sonnet 5.5 and Muse Wonder Full, both at 56 points.

The model's most dramatic gains are evident in agentic tasks and coding capabilities. On the GDPval-AA benchmark, the new version records 1,846 Elo points, a 312-point jump from the previous iteration. This places it just behind Claude Opus 5.5 and Muse Wonder Full. In terminal-based coding tests, specifically Terminal-Bench 4.0, Moonshot reports a success rate of 65.7 percent. This figure surpasses GPT-6 Astra at 58.2 percent and approaches Claude Opus 5.5, which scores 66.4 percent. The older K3 model had managed only 13 percent on this same evaluation.

Hallucination rates have also improved significantly. According to Moonshot’s internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent in the previous model to 8.3 percent in K3.1 Pro. The company notes that this is a stress test rather than a measure of general error rates. For context, GPT-6 Astra achieves a 5.5 percent hallucination rate on the same benchmark, maintaining a slight edge in accuracy.

Beyond general reasoning, the model shows strong performance in 3D modeling and specific professional domains. It can interact directly with game engines like Unity, Unreal Engine, and Godot. On the BenchCAD benchmark, K3.1 Pro scores 92.7 percent when using tools, slightly behind Claude Opus 5.5 at 96.2 percent. Without external tools, it scores 72.8 percent, nearly matching the 73.0 percent achieved by Claude Opus 5.5. Moonshot states the model is particularly effective at parametric modeling of mechanical parts and architectural structures, and it can write scripts for Blender and CAD software without requiring MCP connectors.

The model retains the same architectural dimensions as its predecessor, featuring 2.8 trillion total parameters with 104 billion active ones. It supports a one-million-token context window and handles both text and image inputs. Artificial Analysis places it at the top of the Pareto frontier, meaning no other model with a comparable parameter count achieves a higher intelligence score. The cost per intelligence index task remains roughly stable at $1.98, as the model generates 14 percent more output tokens than the previous version.

Moonshot has also introduced cybersecurity considerations into its training process. The company observed that vulnerability research capabilities grew as post-training scaled up. To prepare for the open-weights release, scheduled for November 4, 2026, Moonshot removed cyber-related data from both training phases. This adjustment reduced the model’s ability to solve recently disclosed vulnerabilities from 90.5 percent to 81.0 percent on an internal benchmark. This is lower than DeepSeek-V4.1-Flash at 85.7 percent, but significantly higher than the previous K3 model at 33.3 percent.

The company aims to balance offensive and defensive cyber capabilities in its open-weights models. Moonshot believes open models should not push the absolute frontier of offensive cyber capabilities but should gradually improve to support defensive tasks like auditing legacy systems. If the weights are released as planned, K3.1 Pro would become the highest-scoring open-weights model available, shrinking the gap between open and proprietary systems from 12 points to just 3.