Moonshot releases Kimi K3.1 Pro with reduced cyber capabilities and lower costs

Moonshot AI introduced Kimi K3.1 Pro, the first update to its K3 family, featuring 2.8 trillion parameters. Benchmarks show the model trails Opus 5.5 by less than a point on Terminal Bench 4.0 and ranks fourth on the Intelligence Index with a score of 55, behind Opus 5.5, Sonnet 5.5, and Muse Wonderful. The model is priced at two-thirds the cost of its predecessor, K3, matching Sonnet-55 and GPT-6 Sol, with a cost of approximately $5 per task compared to $15 for Opus 5.5. Moonshot has also announced plans to release the model weights for free in early November, which would make it the strongest open model available.

The model shows significant improvements in agentic tasks, particularly in command-line operations where it performs nearly on par with Claude Opus 5.5, as well as in office work and financial analysis. It features specialized capabilities for 3D modeling, allowing it to work directly with game engines like Unity, Unreal, and Godot, and to write scripts for Blender and CAD programs. However, Moonshot deliberately reduced the model's offensive cyber capabilities during post-training by removing cyber-related data to keep it below the frontier of proprietary models. Artificial Analysis found K3.1 scored 64% on their cyber index, making it the most capable public cyber model, 11 points ahead of DeepSeek V4.1 Flash. Researchers like Eli Bakaush view this deliberate cyber capability reduction as a positive precedent for open labs.

The launch was marred by an influencer campaign that instructed partners to omit paid partnership labels and trash-talk competitors, drawing criticism from researchers like Nathan Lambert who noted cultural communication issues. Critics argue the undisclosed influencer campaign damaged the launch's credibility, while the technical performance and safety measures have been viewed positively by the broader AI community.