Stomatal frequency (such as stomatal index-SI and stomatal density-SD) has been widely used to reconstruct atmospheric CO2 levels in geological history as it is one of the most reliable proxies of paleo-CO2 that predate the oldest icecore records. However, living plants show large variations on stomatal frequency within the same species, potentially generating large error margins for estimated paleo-CO2 levels using limited fossil specimen(s). The extraordinarily wellpreserved and abundant fossil leaves from the mid-Miocene (~15Ma) Clarkia Lake deposits in northern Idaho, the USA, allow us to test variations within a population of a fossil species and to compare that cross different contemporary species. Our preliminary results from the SD of 15 cuticular membranes of Taxodium revealed a range of variation leading to CO2 levels of 345-445 parts per million (ppm). The SI of eight cuticular membranes of Metasequoia from the same fossiliferous layers reconstructed CO2 levels of 290-345 ppm. These wide and discrepant ranges imply that randomly selected fossil leaves with limited sample numbers may give a large range of CO2 reconstructions and different methods (such as SD or SI) and different plant taxa (such as Taxodium and Metasequoia) may result in different CO2 results. A better understanding of stomatal frequency variations within populations and consistent sampling method will reduce errors in paleo-CO2 reconstruction.