The extent of CO2 transfer from subducting lithologies to the overlying mantle wedge in general and to the arc magma source regions in particular remains debated. The limit of CO2 transfer to the sub-arc mantle could be estimated if the effects of CO2 on the primary hydrous melt compositions of mantle wedge can be assessed in relation to the observed compositions of primitive arc magmas. Here we present new piston cylinder and multi-anvil experiments using Au75Pd25 and Au capsules on four depleted peridotite + H2O ± CO2 starting compositions with 3.5 wt.% H2O and XCO2 [=molar CO2/ (CO2+H2O)] of 0–0.17. Experiments were performed at 2–4 GPa and 1200°C to constrain how the presence of variable CO2 in slab-derived aqueous fluids affects the composition of peridotite partial melts. All experiments consisted of low degree melts (<10 wt.%) in equilibrium with olivine + orthopyroxene ± clinopyroxene. Melts at 2–4 GPa are basaltic for XCO2 of 0–0.10 and become SiO2-poor and CaO-rich at XCO2 > 0.10. Comparison between our experimental partial melt compositions with a global dataset of the most primitive arc magmas suggests that the upper limit of XCO2 in fluids inducing melting in mantle wedges is ∼0.10 at 2–4 GPa. We apply these new constrains to an H2O and CO2 mass balance model for subduction zones and estimate that at least 34–86% of CO2 entering subduction zones bypasses the sub-arc melt generation zone and is subducted to the convecting mantle, either carried by the slab or by the down-dragged limb of the mantle wedge directly above the slab.