Reasoning models don't always say what they think