Air withdrawn from the top 5–15 m of the polar snowpack (firn) shows anomalous enrichment of heavy gases during summer, including inert gases. Following earlier work, we ascribe this to thermal diffusion, the tendency of a gas mixture to separate in a temperature gradient, with heavier molecules migrating toward colder regions. Summer warmth creates a temperature gradient in the top few meters of the firn due to the thermal inertia of the underlying firn and causes gas fractionation by thermal diffusion. Here we explore and quantify this process further in order to (1) correct for bias caused by thermal diffusion in firn air and ice core air isotope records, (2) help calibrate a new technique for measuring temperature change in ice core gas records based on thermal diffusion [ Severinghaus et al., 1998 ], and (3) address whether air in polar snow convects during winter and, if so, whether it creates a rectification of seasonality that could bias the ice core record. We sampled air at 2-m-depth intervals from the top 15 m of the firn at two Antarctic sites, Siple Dome and South Pole, including a winter sampling at the pole. We analyzed ¹⁵N/¹⁴N, ⁴⁰Ar/³⁶Ar, ⁴⁰Ar/³⁸Ar, ¹⁸O/¹⁶O of O₂, O₂/N₂, ⁸⁴Kr/³⁶Ar, and ¹³²Xe/³⁶Ar. The results show the expected pattern of fractionation and match a gas diffusion model based on first principles to within 30%. Although absolute values of thermal diffusion sensitivities cannot be determined from the data with precision, relative values of different gas pairs may. At Siple Dome, δ⁴⁰Ar/4 is 66 ± 2% as sensitive to thermal diffusion as δ¹⁵N, in agreement with laboratory calibration; δ¹⁸O/2 is 83 ± 3%, and δ⁸⁴Kr/48 is 33 ± 3% as sensitive as δ¹⁵N. The corresponding figures for summer South Pole are 64 ± 2%, 81 ± 3%, and 34 ± 3%. Accounting for atmospheric change, the figure for δO₂/N₂/4 is 90 ± 3% at Siple Dome. Winter South Pole shows a strong depletion of heavy gases as expected. However, the data do not fit the model well in the deeper part of the profile and yield a systematic drift with depth in relative thermal diffusion sensitivities (except for Kr, constant at 34 ± 4%), suggesting the action of some other process that is not currently understood. No evidence for wintertime convection or a rectifier effect is seen.