Different stochastic frontier models for panel data are used to estimate education production functions and the efficiency of Finnish general upper secondary schools. Grades in the matriculation examination are used as an output and explained with the comprehensive school grade point average, parental socio-economic background, school resources, the length of studies and the decentralization of test-taking. Heterogeneity across schools is allowed for by estimating true random effect (TRE), random parameter (RP) and true fixed effect (TFE) models. The results show that inefficiency and rankings of schools based on their inefficiency scores vary considerably depending on the type of stochastic frontier model applied. The lowest estimates for inefficiency are obtained with TRE, RP and TFE models, which separate time-constant random or fixed effects from inefficiency. The length of studies and the decentralization of test-taking negatively affect student achievement. (Contains 16 notes, 3 tables, and 2 figures.)