55 Points: Kimi K3.1 Pro Narrows the Gap With Top AI Models

Fifty-five. That is the score Moonshot AI’s new model, Kimi K3.1 Pro, just landed on the Artificial Analysis Intelligence Index. It might not look like a record-shattering number at first glance, but it represents a significant shift in the competitive landscape. This post-trained update to the previous Kimi K3 model has jumped eleven points, vaulting it into fourth place globally.

The only models currently ranking higher are Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full, which sit at 58, 56, and 56 points respectively. This places Kimi K3.1 Pro just three points behind the top proprietary options. If the company follows through on its plan to release the model weights by November 4, 2026, this new version would become the highest-scoring open-weights model available. Such a move would shrink the performance gap between open and proprietary systems from twelve points down to just three, assuming no other major releases happen in the interim.

The most dramatic improvements are seen in agentic tasks, where the model demonstrates a clear leap in capability. On the GDPval-AA benchmark, the new model achieves 1,846 Elo points, a 312-point increase over its predecessor. In terminal-based testing, specifically Terminal-Bench 4.0, Moonshot reports a score of 65.7 percent. This is a massive improvement from the 13 percent recorded by the older K3 model and places it nearly on par with Claude Opus 5.5, which scores 66.4 percent. It also outperforms GPT-6 Astra in this specific category.

Accuracy has also improved, with hallucination rates dropping sharply. According to Moonshot’s internal stress tests, claim-level hallucinations fell from 23.4 percent in the previous version to 8.3 percent. While GPT-6 Astra still holds an edge with a 5.5 percent rate on the same benchmark, Moonshot notes that this metric is designed to stress-test the model rather than measure general error rates. The model also shows strong gains in finance, improving by 7.2 points in financial modeling and 11.6 points in financial research.

Beyond standard text and code, the model has expanded its utility into 3D modeling and game development. It can now interact directly with engines like Unity, Unreal Engine, and Godot to handle everything from quick fixes to full project builds. On the BenchCAD benchmark, it scores 92.7 percent when using tools, a figure close to the top-tier proprietary models. Without tools, it hits 72.8 percent, nearly matching Claude Opus 5.5. Users can also access a new "3D Make" feature on Kimi.com to preview and export models.

Perhaps most intriguing for businesses and developers is the pricing structure. Despite the performance jump, Moonshot has cut API prices by 33 percent. Input costs are now $2 per million tokens, down from $3, while output costs have dropped to $10 from $15. The model retains the same massive scale as its predecessor, with 2.8 trillion total parameters and 104 billion active ones, supporting a context window of one million tokens. A dedicated Highspeed variant is also available, running two to three times faster for those prioritizing latency over raw depth.

Moonshot has also addressed cybersecurity concerns, a topic that often arises with powerful open models. The company stripped cyber-related data from the training process for the open-weights release to balance offensive and defensive capabilities. In internal tests reproducing recent vulnerabilities in software like the Linux kernel, the uncleaned version solved 90.5 percent of tasks, while the final released model solves 81.0 percent. This is comparable to DeepSeek-V4.1-Flash, which scores 85.7 percent, and far above the previous Kimi K3 score of 33.3 percent.

The strategy reflects a broader industry debate about how much offensive cyber capability should be accessible in open models. Moonshot argues that while no open model should push the absolute frontier of offensive capabilities, they should still be capable enough for defensive auditing of legacy systems. The release marks a notable step for Chinese AI companies, suggesting that the gap between homegrown and Western proprietary models is narrowing faster than many expected. For Midwest tech hubs and developers relying on API costs, this price-performance combination is already available for testing.