The Danger of Overthinking in Reasoning Models