Kimi K3 Matches GPT-5.6 Terra on 23 Flaws at One-Quarter the Cost
Updated
Updated · South China Morning Post · Jul 22
Kimi K3 Matches GPT-5.6 Terra on 23 Flaws at One-Quarter the Cost
3 articles · Updated · South China Morning Post · Jul 22
Summary
Kimi K3 found 23 of 26 known vulnerabilities in Aikido’s private benchmark, equaling OpenAI’s mid-tier GPT-5.6 Terra in bug detection.
One-quarter the cost of OpenAI’s flagship Sol model made that result stand out, suggesting Moonshot AI has sharply narrowed the performance gap with top U.S. systems.
Aikido researcher Philippe Dourassov said the test used recently discovered vulnerabilities in a private setup, reducing the chance Kimi K3 had been trained on the answers.
The result is likely to intensify debate over whether U.S. safety curbs are slowing frontier AI progress as Chinese and open-source models gain ground.