Quinn Slack discusses a new metric called "Off-the-Rails Cost," which compares the performance of AI models Sonnet, Gemini, and Opus. He highlights that 17.8% of costs for Gemini users are tied to "wasted threads," significantly worse than the other models. This analysis aims to improve Amp's functionality and may lead to automatic detection of these issues.