Accurate estimation of winter wheat frost kill in cold-temperate agricultural regions is limited by lack of data on soil temperature at wheat crown depth, which determines winter survival. We compared the ability of four models of differing complexity to predict observed soil temperature at 2 cm depth during two winter seasons (2013-14 and 2014-15) at Ultuna, Sweden, and at 1 cm depth at Ilseng and Ås, Norway. Predicted and observed soil temperature at 2 cm depth was then used in FROSTOL model simulations of the frost tolerance of winter wheat at Ultuna. Compared with the observed soil temperature at 2 cm depth, soil temperature was better predicted by detailed models than simpler models for both seasons at Ultuna. The LT50 (temperature at which 50 % of plants die) predictions from FROSTOL model simulations using input from the most detailed soil temperature model agreed better with LT50 FROSTOL outputs from observed soil temperature than what LT50 FROSTOL predictions using temperature from simpler models did. These results highlight the need for simpler temperature prediction tools to be further improved when used to evaluate winter wheat frost kill.