Looking at the code for LogisticRegression.predict_proba it seems to assume that multiclass='ovr': it does sigmoid(wTx+b) and then normalizes over the classes. I think that when multiclass='multinomial' is used it should use softmax instead.