Contemporary exposure therapy models for anxiety argue that exposures must generate threat prediction error to be effective. More research is needed to test this claim in clinical settings. This study explored how threat prediction error learning relates to outcomes during an exposure analogue procedure. Adult undergraduate psychology students (N = 125) experiencing a broad range of social anxiety symptoms from healthy to clinical levels of social anxiety completed 667 online speech performance exposures over two testing sessions separated by a week (approx. 3 speeches/session). Self-reported anxiety, threat prediction, threat outcome, and surprise were measured for each exposure and used to derive learning indicators. These included threat prediction error, prediction change, and the extent that prediction errors were converted to prediction change (i.e., learning rate). We examined between- and within-person relationships between these learning indicators and outcomes over exposure using multilevel modelling. Average prediction change and prediction error learning rate, but not average prediction error per se, was associated with more anxiety reduction across the exposure. Within-person, anxiety was lower after exposures that triggered more prediction change. Threat prediction error was not linearly associated with anxiety at the next exposure. Higher threat prediction error during an exposure was associated with greater subjective surprise for that exposure. We concluded that exposure outcomes depend on how much the patient converts exposure-related prediction errors into threat prediction change. Future research should focus on strategies to enhance the prediction-error learning rate from exposures.