China’s Kimi K3.1 Pro closes gap with top closed AI models

Key Points

1. Moonshot AI released Kimi K3.1 Pro on October 23, 2026, with open weights due by November 4.
2. Independent testing places it fourth on intelligence, ahead of Claude Sonnet 5.5 on terminal coding tasks.
3. Pricing at $2 input per million tokens undercuts closed rivals, widening enterprise options in the region.

The latest:

A Chinese open-weight model is now within striking distance of the strongest closed systems on coding work. Moonshot AI launched Kimi K3.1 Pro and a low-latency Highspeed variant on October 23, 2026, a mixture-of-experts system with 2.8 trillion total parameters and 104 billion active per token. Artificial Analysis expects it to top the open-weight field once weights ship.

Details:

• The architecture: Moonshot describes K3.1 Pro as a post-trained version of Kimi K3 built on substantially more post-training compute rather than a new architecture. It carries a 1 million token context window and accepts both text and images. The company names coding, 3D modeling and finance as its focus domains.
• Company benchmarks: Moonshot reported 65.7% on Terminal-Bench 4.0 using Claude Code, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. On BenchCAD for 3D modeling it posted 92.7%, below Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.
• Financial documents: On OfficeQA Pro v2, which tests handling of financial documents, Moonshot reported 64.4 and described the result as best in class. It did not publish competitor figures for that benchmark alongside its own claim.
• Independent testing: Artificial Analysis scored the model at 55.1 on its intelligence index, fourth overall and a sharp rise from 43.6 for K3. Claude Opus 5.5 leads at 57.6. On Terminal-Bench 4.0 the independent run put K3.1 Pro at 67%, ahead of Claude Sonnet 5.5 at 64%.
• Hallucination rate: On an internal hard-case benchmark, Moonshot said the hallucination rate fell from 23.4% to 8.3%, a 65% reduction. The company cautioned those figures do not represent error rates in everyday use.
• The pricing: K3.1 Pro costs $2 per million input tokens, $10 per million output tokens and $0.20 for cached input. The Highspeed variant runs $4 and $20 with the same cached-input price, and Moonshot says it delivers 2 to 3 times the speed.
• Availability: The model is live through Kimi.com, the Kimi apps, Kimi Work, Kimi Code and the Kimi API. Full weights are scheduled for release by November 4, 2026, and have not yet been published.
• Cybersecurity curbs: Moonshot said it stripped cybersecurity-related training data ahead of the open release. An unrestricted version solved 90.5% of tasks on an internal exploitation test while the shipped model solved 81.0%, with the company stating its aim is not to extend the offensive capability frontier.
• The projection: Artificial Analysis expects K3.1 Pro to become the highest-scoring open-weight model once released, narrowing the distance to closed frontier systems from 12 points to 3 on its index.

Background:

Open-weight models publish their parameters for anyone to download, run locally and modify, unlike closed systems reachable only through a vendor’s API. That distinction governs where enterprises can deploy a model and what data they are willing to send to it.

Between the lines:

The gap Moonshot is attacking is price, not peak capability. Its own benchmarks place it just below Opus 5.5 on coding and clearly below on 3D modeling, yet input pricing at $2 per million tokens reframes the comparison for buyers running high volume. The cybersecurity data removal signals a company anticipating scrutiny of what downloadable weights enable — a concern that does not apply to API-only rivals.

What’s next

The November 4, 2026 weight release is the decisive marker: whether Moonshot ships on schedule, whether independent labs reproduce the 55.1 intelligence score, and whether anyone verifies the exploitation-test figures once the weights are public.