Confidence intervals for the median lethal dose (LD50) and other dose percentiles in logistic regression models are developed using a generalization of the Fieller theorem for exponential families and saddlepoint approximations. Simulation results show that, in terms of one-tailed and two-tailed coverage, the proposed methodology generally outperforms competing confidence intervals obtained from the classical Fieller, likelihood ratio, and score methods. In terms of two-tailed coverage, the proposed method is comparable to the Bartlett-corrected likelihood ratio method, but generally outperforms it in terms of one-tailed coverage. An extension to the competing risk setting is presented that allows binary response adjustments to be made using observed censoring times. Supplementary materials for this article are available online at http://pubs.amstat.org.