AbstractThe sensitivity of coccolithophores to changing CO2 and its role modulating cellular photosynthetic carbon isotopic fractionation (εp) is crucial to understand the future adaptation of these organisms to higher CO2 world and to assess the reliability of εp for past CO2 estimation. Here, we present εp measured on natural fossil samples across the glacial‐interglacial (G‐I) CO2 variations of marine isotope stages 12 to 9 interval (454–334 ka) at the western tropical Atlantic Ocean Drilling Program Site 925 together with a set of organic and inorganic geochemical, micropaleontological and morphometrical data from Gephyrocapsa coccoliths in the same samples. The ∼2‰ variation in εp is significantly correlated with the CO2[aq] concentrations calculated from assumption of air‐sea equilibrium with measured ice core pCO2 concentrations. The sensitivity of εp to CO2[aq] is similar to that derived from a multiple regression model of culture observations and is not well simulated with the classical purely diffusive model of algal CO2 acquisition. The measured range of Gephyrocapsa cell sizes is insufficient to explain the non‐CO2 effects on εp at this location, via either direct size effect or growth rate correlated to cell size. Primary productivity, potentially triggered by shifting growth rates and light levels, may also affect εp. Proposed productivity proxies % Florisphaera profunda and the ratio between the C37 to C38.et alkenone (C37/C38.et ratio) both correlates modestly with the non‐CO2 effects on εp. When the observed G‐I εp to CO2 sensitivity at this site is used to estimate pCO2 from εp since the Miocene, the inferred pCO2 declines are larger in amplitude compared to that calculated from a theoretical εp diffusive model. We find that oxygen and carbon stable isotope vital effects in the near monogeneric‐separated Gephyrocapsa coccoliths (respectively Δδ18OGephyrocapsa–Trilobatus sacculifer and εcoccolith) are coupled through the time series, but the origins of these vital effects are not readily explained by existing models.