Kimi K3.1 Pro: China’s Open-Weight Model Beats GPT-6 and Gemini 4
Trending Topics Jakob Steinschaden ● Covered by 4 sources
Moonshot’s Kimi K3.1 Pro has jumped ahead of GPT-6 Astra and Gemini 4 on a major A.I. index. If its weights are released in November, it could become the strongest open model by a wide margin.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Moonshot AI has put out Kimi K3.1 Pro, and on Artificial Analysis’s Intelligence Index it lands at 55 points. That puts it fourth overall, just behind Anthropic’s Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Meta’s Muse Wonder Full at 56. OpenAI’s GPT-6 Astra and Google’s Gemini 4 Argon are both on 53. The jump from Kimi K3 is sizeable too: 44 points to 55.
The bigger shift may come later. Moonshot says it plans to release the model’s weights by early November. If that happens, Artificial Analysis says Kimi K3.1 Pro would become the world’s strongest open-weight model, and the gap between the best open model and the best proprietary one would shrink from 12 points to 3. Among open models, that would move Kimi well ahead of Xiaomi’s MiMo-V2.6-Pro, DeepSeek V4.1 Flash and its own predecessor.
The model’s clearest gains show up in agentic work, the kind of thing that involves multiple steps and more than one tool. On GDPval-AA, which measures realistic knowledge work, Kimi K3.1 Pro scores 1,846 Elo points, up 312 from K3. That places it behind Claude Opus 5.5 and Muse Wonder Full, but it is a meaningful leap. On Terminal-Bench 4.0, it hits 67 percent, up from 13 percent on K3, and above Claude Sonnet 5.5’s 64 percent. On AutomationBench-AA, the score rises from 58 to 73 percent.
Moonshot says the underlying architecture barely changed. It still has 2.8 trillion parameters in total, with 104 billion active per request, plus a one-million-token context window and support for text and images. The company did, however, pour more compute into post-training. Pricing is lower than before, with API rates at $2 per million input tokens and $10 per million output tokens, while the cost per task on the index stays about the same because the model generates more text.
There is also a Highspeed version that Moonshot says is two to three times faster and costs twice as much. And Moonshot has trimmed cyberattack-related training data before the open-weight release, which the company says reduces offensive capability. That trade-off is the whole open-model story in a nutshell: more access, more scrutiny, and a lot less control once the weights are out in the wild.
My take — AI-written commentary, not fact-checked reporting
Open weights keep winning the public relations war because they can be copied, shared and benchmarked, which is very convenient for people who love charts. The catch is the same one every time: once those weights are out, nobody gets them back, and the safety debate stops being theoretical. The industry keeps acting surprised that openness and control don’t like each other much.
Read more about this at: Trending Topics