“Based on internal experimentation, Anthropic has found that for complex tasks like bug hunting, using a "best-of-N" strategy provides more performance returns than simply using a bigger model or increasing runtime.”