OBJECTIVE To evaluate the predictive validity of critical power (CP) and the work above CP (W') on cycling performance (mean power during a 20-min time trial; TT20). METHODS On 3 separate days, 10 male cyclists completed a TT20 and 3 CP and W' prediction trials of 1, 4, and 10 min and 2, 7, and 12 min in field conditions. CP and W' were modeled across combinations of these prediction trials with the hyperbolic, linear work/time, and linear power inverse-time (INV) models. The agreement and the uncertainty between the predicted and actual TT20 were assessed with 95% limits of agreement and a probabilistic approach, respectively. RESULTS Differences between the predicted and actual TT20 were "trivial" for most of the models if the 1-min trial was not included. Including the 1-min trial in the INV and linear work/time models "possibly" to "very likely" overestimated TT20. The INV model provided the smallest total error (ie, best individual fit; 6%) for all cyclists (305 [33] W; 19.6 [3.6] kJ). TT20 predicted from the best individual fit-derived CP, and W' was strongly correlated with actual TT20 (317 [33] W; r = .975; P < .001). The bias and 95% limits of agreement were 4 (7) W (-11 to 19 W). CONCLUSIONS Field-derived CP and W' accurately predicted cycling performance in the field. The INV model was most accurate to predict TT20 (1.3% [2.4%]). Adding a 1-min-prediction trial resulted in large total errors, so it should not be included in the models.