Moonshot AI's Kimi K3.1 Pro Claims Fourth Place in Global AI Rankings

Moonshot AI released Kimi K3.1 Pro, a post-trained reasoning model that claims the number four spot on the Artificial Analysis Intelligence Index. The model scored 55 points, an 11-point increase over its predecessor, Kimi K3. This result places it just three points behind the top proprietary models currently available on the market.

According to the Artificial Analysis index, three models currently rank higher than K3.1 Pro: Claude Opus 5.5, which holds the top position with 58 points, followed by Claude Sonnet 5.5 and Muse Wonder Full, which both posted 56 points. Moonshot has scheduled a release of model weights by November 4, 2026. If the release proceeds as planned, K3.1 Pro would become the highest-scoring open-weights model, reducing the gap between open and proprietary models to just 3 points.

The model's strongest performance gains appear in agentic tasks. On the GDPval-AA benchmark, K3.1 Pro reached an Elo of 1,846, an improvement of 312 points over K3. On Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent, which trails Claude Opus 5.5's 66.4 percent but exceeds GPT-6 Astra's 58.2 percent. In comparison, the K3 model scored only 13 percent on the same test. Moonshot also reports that claim-level hallucinations dropped from 23.4 percent in K3 to 8.3 percent in K3.1 Pro on its internal Kimi Hallu Bench Hard stress test.

K3.1 Pro maintains the same parameters as its predecessor, featuring 2.8 trillion total parameters and 104 billion active parameters, with a context window spanning one million tokens. While the size remains constant, API prices have dropped by 33 percent. Input costs $2 per million tokens, down from $3, while output costs $10 per million tokens, down from $15. The model is available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. A variant called K3.1 Pro Highspeed is also shipping, which Moonshot reports runs two to three times faster.

Moonshot has also implemented specific limitations regarding cybersecurity capabilities for the open-weights release. The company stripped cyber-related data from both training phases to prevent the model from pushing the absolute frontier of offensive cyber capabilities. On the internal xdev-bench-hard test, the uncleaned version of the model solves 90.5 percent of tasks, while the data-cleaned K3.1 Pro solves 81.0 percent. The company states that while it wants to raise what open models can do for defensive purposes like auditing legacy systems, it does not want them to push the absolute frontier of offensive capabilities.