Chinese AI model closes gap with top rivals

The gap between leading proprietary artificial intelligence systems and new Chinese entrants has narrowed to just three points. This shift marks a significant milestone in the global race for advanced reasoning capabilities, challenging the previous dominance of established Western tech giants.

Moonshot AI released Kimi K3.1 Pro, which scored 55 points on the Artificial Analysis Intelligence Index. This result represents an 11-point improvement over its predecessor, Kimi K3. Only three models currently outperform it: Anthropic's Claude Opus 5.5 with 58 points, Claude Sonnet 5.5 with 56 points, and Muse Wonder Full with 56 points. If Moonshot publishes the model weights by November 4, 2026, as planned, K3.1 Pro would become the highest-scoring open-weights model available.

The company is also offering a variant called K3.1 Pro Highspeed, which runs two to three times faster than the standard version. Both models are accessible through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The primary architectural changes involve substantial increases in post-training compute rather than structural modifications to the base model.

Performance gains were most evident in agentic tasks. On the GDPval-AA benchmark, the new model reached 1,846 Elo, a 312-point increase over K3. This places it slightly behind Claude Opus 5.5 and Muse Wonder Full. In Terminal-Bench 4.0 testing, Moonshot reported a score of 65.7 percent, surpassing GPT-6 Astra at 58.2 percent and trailing only Claude Opus 5.5 at 66.4 percent. The previous K3 model had managed only 13 percent on that specific test.

The update also addresses common AI errors. Internal testing shows a sharp decline in hallucinations, dropping from 23.4 percent in K3 to 8.3 percent in K3.1 Pro. While GPT-6 Astra still leads in this metric with a 5.5 percent rate, Moonshot describes its test as a stress test for difficult claims rather than a measure of general error rates. Financial modeling and research capabilities also improved, with gains of 7.2 and 11.6 points respectively.

Capabilities in 3D modeling and coding have expanded. The model can now write and run scripts for Blender and CAD software without additional connectors and works with game engines like Unity and Unreal. On the BenchCAD benchmark, it achieved 92.7 percent accuracy with tools, compared to 96.2 percent for Claude Opus 5.5. Without tools, it reached 72.8 percent, nearly matching the 73.0 percent recorded by the top rival. A new 3D Make feature allows users to preview and export models directly on Kimi.com.

The underlying architecture remains consistent with K3, featuring 2.8 trillion total parameters and 104 billion active parameters. It supports a one-million-token context window for text and images. API prices have dropped by 33 percent, with input costs at $2 per million tokens and output at $10. Despite generating 14 percent more output tokens, the total cost per Intelligence Index task is $1.98, nearly identical to the previous model's $2.00 cost.

Cybersecurity capabilities grew as post-training scaled, with the uncleaned model solving 90.5 percent of tasks on an internal benchmark that reproduces recent vulnerabilities in software like the Linux kernel. After removing cyber-related data for the open-weights release, the score adjusted to 81.0 percent. Moonshot notes that while no open-weights model should push the absolute frontier of offensive cyber capabilities, gradual improvements support essential defensive work such as auditing legacy systems.