Ramp Labs finds GPT-5.5 beats Kimi K2.6

An analysis of Ramp Labs' security scanning data reveals that while cheaper AI models like Kimi K2.6 can outperform frontier models such as GPT-5.5 on a small budget due to higher volume, they hit a performance ceiling on subtle bugs. The author demonstrates that once a budget allows for three frontier model tries per bug, GPT-5.5 surpasses the detection rates of the cheaper models. A script is provided to help teams calculate pass@k probabilities for their own model swarms. Open article →