A new framework for estimating confidence in large language models, called Divergent Token Confidence (DTC), is proposed. DTC estimates confidence by counting tokens at which two models strongly disagree during decoding, and is shown to improve calibration over probability-based methods.