For most of the Phanerozoic Eon, Earth’s woody vegetation has been dominated by C3 plants - predominantly gymnosperms - with angiosperms only emerging as the dominant plant group as CO2 declined during the Cenozoic (66 Ma onward). At present, differences in carbon isotope discrimination (Δ13C) between angiosperm and gymnosperm plants are relatively small (2–3 ‰), but an increasing body of evidence points to larger differences across geological times (up to 6–7 ‰), potentially associated with varying environmental conditions and atmospheres (i.e. concentrations of atmospheric carbon dioxide, [CO2], and oxygen, [O2] could have ranged from ∼180 to 1100 ppm, and ∼15 to 25%, respectively, across the past 250 Ma). Yet, differences in Δ13C between the two plant groups, and their potential link to climatic and environmental changes, have not yet been fully explored and understood. Here, we combine a comprehensive ab initio model of discrimination, with a recent model of plant eco-physiology based on least-cost optimality theory, to show how differences in Δ13C between angiosperms and gymnosperms arise. We train the comprehensive model using a very large (n>7000) database of leaf and tree ring data spanning the past 110 years. We find that averaged differences in Δ13C between angiosperm and gymnosperms decrease modestly with atmospheric [O2]:[CO2] ratios, and increase strongly with vapor pressure deficit (D). These relationships can be explained by three key physiological differences: (1) the ratio of cost factors for transpiration to carboxylation (higher in angiosperms); (2) the ratio of mesophyll to stomatal conductances of CO2 (lower in gymnosperms); and (3) differences in photorespiration. In particular, the amount of CO2 released from photorespiration per oxygenation reaction, λ, is generally lower in gymnosperms than in angiosperms. As a result of these factors, Δ13C is more sensitive to [CO2] in angiosperms, and to D in gymnosperms. We propose a simplified empirical model to account for this behaviour, and test it against isotopic data from leaves, tree rings and previously-published plant chamber experiments, along with geological data from the Cenozoic. Overall, these data agree with our model over a range of [O2]:[CO2] ratios from 100 to 650 mol mol−1 (equivalent to a CO2 range around 323–2100 ppm at 21% O2), and D levels between 0.45 and 1.1 kPa (R2 = 0.51, RMSE = 1.49 ‰). Our simplified empirical model offers a new explanation for secular trends in the geological record, and suggests a way forward to improve paleo-[CO2] proxies based on terrestrial discrimination models by incorporating the effects of [O2], phylogeny, and photorespiration. Lastly, the framework predicts that the average difference in Δ13C between woody C3 plant groups will increase in the future if both [CO2] and global D continue to rise as suggested by projections.