AI Race: Moonshot's New Model Narrows Gap with Anthropic

Fifty-five points. That is the score that recently propelled a new Chinese AI model into the upper echelon of global artificial intelligence rankings, marking a significant shift in the competitive landscape.

Moonshot AI has officially released Kimi K3.1 Pro, a post-trained iteration of its predecessor, Kimi K3. The update represents a substantial leap in performance, registering an 11-point increase over the previous version. This boost secures the model a fourth-place finish on the Artificial Analysis Intelligence Index, placing it in direct contention with some of the most advanced proprietary systems currently available.

The numbers paint a picture of rapid convergence. Only three models currently outperform Kimi K3.1 Pro: Claude Opus 5.5 with a score of 58, Claude Sonnet 5.5 with 56, and Muse Wonder Full with 56. The margin between the new Chinese model and the top proprietary entries is just three points. If Moonshot adheres to its plan to release the model weights by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model on the market. Such a move would drastically reduce the performance gap between open and proprietary architectures from twelve points to merely three, assuming no other major releases intervene.

Alongside the primary release, the company is shipping a variant named K3.1 Pro Highspeed. According to Moonshot, this version operates two to three times faster than its standard counterpart. Both versions are immediately accessible through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, allowing developers and businesses to integrate the new capabilities right away.

The most dramatic improvements in K3.1 Pro are visible in agentic tasks, which involve the model autonomously executing complex workflows. On the GDPval-AA benchmark, the new model achieves an Elo rating of 1,846, a massive 312-point surge over the previous K3 version. It now trails only Claude Opus 5.5 and Muse Wonder Full in this specific metric. Performance on Terminal-Bench 4.0 also saw a dramatic uptick, with Moonshot reporting a score of 65.7 percent. This places the model ahead of GPT-6 Astra and just barely behind Claude Opus 5.5, while the older K3 model had managed only 13 percent on the same test.

Reliability has also improved, with a sharp decline in hallucination rates. On Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucinations dropped from 23.4 percent in the previous version to 8.3 percent. While the company notes this is a stress test rather than a measure of general error rates, the reduction is significant. GPT-6 Astra still holds an advantage with a 5.5 percent rate on the same benchmark. In financial applications, K3.1 Pro showed gains of 7.2 points in modeling and 11.6 points in research, and it currently leads all tested models on the OfficeQA Pro v2 benchmark for general office tasks.

The model demonstrates particular strength in 3D modeling and coding, capable of working directly with game engines such as Unity, Unreal Engine, and Godot. On the BenchCAD benchmark, it scores 92.7 percent when using tools, closely following Claude Opus 5.5 and GPT-6 Astra. Without tools, it achieves 72.8 percent, nearly matching the top proprietary model. Moonshot highlights its ability to handle parametric modeling of mechanical and architectural parts, and it has introduced a '3D Make' feature on Kimi.com that allows users to preview, edit, and export 3D models.

Despite the performance jump, the model maintains the same architectural footprint as its predecessor, featuring 2.8 trillion total parameters and 104 billion active ones. The context window remains at one million tokens, supporting both text and image inputs. However, the cost of using the model has dropped significantly. Token prices have fallen by 33 percent, with input costs now at $2 per million tokens, output at $10, and cache hits at $0.20. This pricing strategy keeps the cost per Intelligence Index task at $1.98, almost identical to the previous version, despite the model generating 14 percent more output tokens.

Moonshot has also placed a heavy emphasis on cybersecurity in its technical documentation. The company observed that vulnerability research capabilities grew as post-training scaled, so it stripped cyber-related data from both training phases for the open-weights release. On its internal xdev-bench-hard test, which reproduces recent software vulnerabilities, the cleaned K3.1 Pro solves 81.0 percent of tasks, down from 90.5 percent in the uncleaned version. This is still well above the 33.3 percent scored by the older K3 model, though below DeepSeek-V4.1-Flash’s 85.7 percent. The company aims to balance offensive capabilities with defensive utility, noting that even lower-scoring open models can write exploits at near-frontier levels.