Moonshot AI Scores 55 Points, Ranking Fourth on Global Intelligence Index
The new Kimi K3.1 Pro model cuts the performance gap with top proprietary systems to three points while reducing API costs by a third.
By Amir Shah ·

Moonshot AI has released Kimi K3.1 Pro, a reasoning model that scored 55 points on the Artificial Analysis Intelligence Index. This result places the Chinese company’s new system in fourth position globally. The score marks an 11-point increase over the predecessor model, Kimi K3.
Three points separate K3.1 Pro from the leading proprietary systems. Anthropic’s Claude Opus 5.5 holds the top spot with 58 points. Claude Sonnet 5.5 and Muse Wonder Full each score 56 points. Moonshot plans to publish the model weights by November 4, 2026. If released as scheduled, K3.1 Pro would become the highest-scoring open-weights model available.
The model shows its largest gains in agentic tasks. On the GDPval-AA benchmark, K3.1 Pro reached 1,846 Elo. This is 312 points higher than the previous version. Claude Opus 5.5 leads that specific test with 1,866 Elo. On Terminal-Bench 4.0, the new model recorded a 65.7 percent score. This exceeds GPT-6 Astra, which recorded 58.2 percent, and trails Claude Opus 5.5’s 66.4 percent by less than one point.
Moonshot reports a sharp decline in hallucinations for the new release. On the internal Kimi Hallu Bench Hard, claim-level hallucinations dropped to 8.3 percent from 23.4 percent in the prior model. GPT-6 Astra still leads that internal test with a 5.5 percent error rate. The company also cites improvements in finance, with an 11.6 point gain on Financial Research benchmarks.
The underlying architecture remains identical to K3. The system features 2.8 trillion total parameters, 104 billion active parameters, and a one-million-token context window. API prices decrease by 33 percent. Input tokens cost $2 per million, while output tokens cost $10. Cache hits are priced at $0.20 per million tokens.
Moonshot adjusted the training data for the open-weights version to address cybersecurity concerns. The company removed cyber-related data from both training phases. On the internal xdev-bench-hard test, the uncleaned model solved 90.5 percent of tasks. The cleaned version solved 81.0 percent. DeepSeek-V4.1-Flash scored 85.7 percent on the same benchmark.
Both the standard K3.1 Pro and a faster Highspeed variant are available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The Highspeed variant operates two to three times faster according to the company.





