Moonshot cuts AI costs by a third

Moonshot AI released Kimi K3.1 Pro on Thursday.

The new reasoning model scores 55 on the Artificial Analysis Intelligence Index. That is an 11 point jump from its predecessor. The score places the system fourth globally. Only Anthropic's Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full rank higher. The gap between the new model and the top tier is now three points.

Artificial Analysis notes that Moonshot plans to publish the model weights by November 4. If the release happens as scheduled, K3.1 Pro becomes the highest scoring open weights model. That would narrow the performance gap between open and proprietary systems from 12 points to 3. The architecture remains identical to the previous version. It features 2.8 trillion total parameters, with 104 billion active. The context window supports one million tokens.

Price adjustments accompany the release. Input costs drop to $2 per million tokens. Output prices stand at $10 per million tokens. Cache hits cost $0.20 per million tokens. These figures represent a 33 percent reduction from previous rates. The model generates 14 percent more output tokens than the prior version. The total cost to run each Intelligence Index task is $1.98. This is nearly identical to the $2.00 cost for the older model.

Performance gains are most visible in agentic tasks. The model reached 1,846 Elo on the GDPval-AA benchmark. That score exceeds the previous version by 312 points. It trails Claude Opus 5.5 by 20 points. On Terminal-Bench 4.0, K3.1 Pro achieves a score of 65.7 percent. GPT-6 Astra scores 58.2 percent on that test. Claude Opus 5.5 leads at 66.4 percent. The older K3 model managed only 13 percent.

Moonshot reports a sharp decline in hallucinations. Internal tests show claim level hallucinations falling from 23.4 percent to 8.3 percent. The company describes this metric as a stress test rather than a general error measure. GPT-6 Astra still leads that specific internal benchmark at 5.5 percent. Financial modeling scores improved by 7.2 points. Financial research scores rose by 11.6 points. Moonshot states the model tops all tested systems on OfficeQA Pro v2.

Technical capabilities in 3D modeling and coding also expanded. The system works with Unity, Unreal Engine, and Godot. It can write scripts for Blender and CAD software without MCP connectors. BenchCAD results show a score of 92.7 percent with tools. That compares to 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra. Without tools, the score is 72.8 percent. Claude Opus 5.5 scores 73.0 percent in the same condition.

Cybersecurity features drew specific attention in the technical documentation. Vulnerability research capabilities improved as post training compute increased. For the open weights release, Moonshot removed cyber related data from both training phases. An internal benchmark called xdev-bench-hard tests the ability to reproduce disclosed vulnerabilities. The uncleaned model solved 90.5 percent of tasks. The cleaned version solved 81.0 percent. DeepSeek-V4.1-Flash scored 85.7 percent. The older Kimi K3 model scored 33.3 percent.