Moonshot AI Cuts Prices While Closing the Gap on Top US Rivals
A new Chinese reasoning model has surged up the intelligence rankings by dropping costs significantly and improving performance in complex technical tasks.
By Andrew Fletcher ·
Moonshot AI has officially launched Kimi K3.1 Pro, a significant update to its existing reasoning model that immediately climbed to the fourth position on the Artificial Analysis Intelligence Index. The new release scores 55 points, placing it just three points behind the current market leaders. This jump represents an 11-point increase over its predecessor, Kimi K3, marking a notable acceleration in the company’s progress against established Western competitors.
The model is available immediately through the company's various platforms, including Kimi.ai, Kimi Work, and its API. Alongside the standard version, Moonshot is also offering a Highspeed variant that operates two to three times faster. If the company follows through on its plan to release the model weights by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model available. This move would reduce the performance gap between open and proprietary models from 12 points to just three, assuming no other major releases occur in the interim.
The most substantial improvements are seen in agentic tasks, where the model demonstrates a clear leap in capability. On the GDPval-AA benchmark, the new model achieved an Elo rating of 1,846, which is 312 points higher than the previous version. It now trails only Claude Opus 5.5 and Muse Wonder Full in this specific category. In coding evaluations, the model reported a 65.7 percent success rate on Terminal-Bench 4.0, a dramatic improvement from the 13 percent recorded by its predecessor. This places it nearly on par with the top proprietary options.
Reliability has also improved, with a sharp decline in hallucination rates. According to internal testing, claim-level hallucinations dropped from 23.4 percent to 8.3 percent. While Moonshot notes this is a stress test rather than a measure of general error rates, the reduction is significant. The model also showed gains in financial modeling and research, and it currently tops all tested models for general office tasks on the OfficeQA Pro v2 benchmark.
In the realm of 3D modeling and game development, the model can interact directly with engines like Unity, Unreal Engine, and Godot. It achieved a 92.7 percent score on the BenchCAD benchmark when using tools, which is close to the scores of top US competitors. Moonshot highlights the model's strength in parametric modeling for mechanical parts and architecture, noting it can write and run scripts for Blender and CAD software without needing additional connectors. A new feature called '3D Make' allows users to preview and edit these models directly on the website.
Despite the performance gains, the model retains the same architecture as its predecessor, with 2.8 trillion total parameters and 104 billion active parameters. However, the cost of using the service has decreased significantly. Input costs have fallen to $2 per million tokens, output costs to $10, and cache hits to $0.20, representing a 33 percent price drop overall. The context window remains at one million tokens, and the model handles both text and image inputs.
Moonshot also addressed cybersecurity concerns by stripping cyber-related data from the training process for the open-weights version. This decision was made to balance offensive capabilities with defensive utility. The cleaned model solves 81.0 percent of tasks on a difficult internal benchmark, compared to 90.5 percent for the uncleaned version. The company believes this approach allows open models to serve defensive purposes, such as auditing legacy systems, without pushing the absolute frontier of offensive cyber capabilities.
The launch positions the model at the top of the Pareto frontier for its size, meaning no model with a comparable parameter count scores higher on the intelligence index. While the cost per task remains nearly identical to the previous version due to slightly increased output token generation, the overall efficiency and capability improvements make it a strong contender for developers and businesses looking for high-performance AI solutions.
Source: The Decoder
