Recent progress of including lake subroutines in numerical weather prediction (NWP) models has led to more accurate forecasts. In lake models, one essential parameter is water clarity, parameterized via the light extinction coefficient, K_d, for which a global constant value is usually used. We used direct eddy covariance fluxes and basic meteorological measurements coupled with lake water temperature and clarity measurements from a boreal lake to estimate the performance of two lake models, LAKE and FLake. These models represent two 1‐D modeling frameworks broadly used in NWP. The results show that the lake models are very sensitive to changes in K_d when it is lower than 0.5 m⁻¹. The progress of thermal stratification depended strongly on K_d. In dark‐water simulations the mixed layer was shallower, longwave and turbulent heat losses higher, and therefore the average water column temperatures lower than in clear‐water simulations. Thus, changes in water clarity can also affect the onset of ice cover. The more complex LAKE modeled the seasonal thermocline deepening, whereas it remained virtually constant during summer in the FLake model. Both models overestimated the surface water temperatures by about 1°C and latent heat flux by >30%, but the variations in heat storage and sensible heat flux were adequately simulated. Our results suggest that, at least for humic lakes, a lake‐specific, but not time‐depending, constant value for K_d can be used and that a global mapping of K_d would be most beneficial in regions with relatively clear lakes, e.g., in lakes at high altitudes.