Challenging molecular separation tasks continue to drive the need for advanced membranes, which could combine high productivity and selectivity with outstanding chemical resistance and mechanical robustness. Here, ultra-thin, sub-100 nm carbon molecular sieve (CMS) membranes based on a polymer of intrinsic microporosity precursor were fabricated and characterized with respect to their separation performance for six gases, He, H2, N2, O2, CO2 and CH4. The rarely reported in thin CMS membranes physical aging was tracked until near-equilibrium was reached. The use of commercially available, small pore size (2–5 nm) γ-alumina-coated α-alumina ceramic supports allowed for an easy membrane precursor fabrication by a simple solution coating process. The subsequent application of a protective polydimethylsiloxane layer, as often done in research and membrane production, assured excellent defect control and produced CMS composite membranes with very high membrane permselectivities (e.g. CO2/CH4 = 84.5, H2/CH4 = 360). No significant deviations from bulk properties in terms of the chemical decomposition were discovered in the film thickness range of 60–300 nm. However, a surprisingly massive acceleration of the physical aging in comparison to the bulk was detected in all sub-100 nm membranes including the precursor polymer. All CMS membranes reached near-equilibrium gas transport properties within 1–1.5 months providing a rare opportunity to study aging-stabilized performance. For the aging-stabilized sub-100 nm membranes a decrease in gas permeabilities of up to 3 orders of magnitude was detected. Remarkably, sub-100 nm CMS membranes showed significantly increased selectivities typical to thick isotropic CMS membranes pyrolyzed at 100–200 °C higher temperatures. A clear aging-related shift to more significant size sieving behavior occurred for the majority of the studied CMS membranes throughout the aging period. Overall, our study provides an important observation that excessive reduction of the selective layer thickness, especially in initially highly microporous materials (PIMs, CMS etc.), may not present the best strategy for the optimization of the long-term membrane performance.