Santorini caldera has had a long history of plinian eruptions and caldera collapses, separated by20–40 kyr interplinian periods. We have carried out a study to constrain magma storage/extractiondepths beneath the caldera. We analysed H2O in 138 olivine-, pyroxene- and plagioclase-hostedmelt inclusions from plinian and interplinian products from the last 200 kyr, and CO2, S, Cl, F anddD in various subsets of these. The dataset includes 64 inclusions in products of the Minoan plinianeruption of the late 17th century BCE. All the melt inclusions were ellipsoidal and isolated, with notextural evidence for volatile leakage. Mafic melt inclusions contain 1–4 wt % H2O and up to1200ppm CO2, 1200ppm S, 2000ppm Cl and 400ppm F; silicic inclusions contain 2–7 wt % H2O, upto 150ppm CO2, up to 400ppm S, 2000–6000ppm Cl and 600–1000ppm F. The dD values of 27 representativeinclusions (–37 to –104%) are intermediate between mantle and slab values and ruleout significant H2O loss by hydrogen diffusion from olivine-hosted inclusions. H2O, S and Cl behavecompatibly in melt inclusion suites varying from mafic to silicic in composition, showing thatentrapment of many melt inclusions took place under volatile-saturated conditions. Most Santorinimelts are saturated in a free COHSCl vapour phase at depths of less than 10 km; the only exceptionsare basaltic melts from a single interplinian eruption, which were volatile-undersaturated upto K2O contents of 1 wt %. The rhyolitic melt of the Minoan eruption probably contained a freehypersaline liquid phase. H2OþCO2 saturation pressures were calculated using suitably calibratedsolubility models to estimate pre-eruptive magma storage depths. Magmas feeding plinian eruptionswere stored at >4km (>100 MPa) and extracted over depth intervals of several kilometres.Plagioclase phenocrysts in rhyodacitic pumice from the Minoan eruption have cores containingmelt inclusions trapped at depths up to 10–12km (320 MPa), and rims (also orthopyroxene andclinopyroxene) containing inclusions trapped at 4–6km (100–160 MPa). This records late-stage silicicreplenishment of a <2km thick shallow magma chamber, rather than extraction of melts syneruptivelyover the entire depth range. The plagioclase cores were carried from depth in the ascendingmelt, then overgrown by the rims in the shallow chamber. Exsolution of volatiles duringascent may have caused the replenishment melt to inject as a bubbly plume, causing mixing priorto eruption. This would explain (1) the homogeneity of the Minoan rhyodacitic magma, and (2) extractionof melt inclusions from the entire pressure spectrum during the first eruptive phase. Mostsilicic magmas feeding eruptions of the interplinian periods were stored in reservoirs at shallowdepths (2–3 km) compared with those feeding the plinian eruptions (>4 km). Melt inclusions fromthe AD 726 eruption of Kameni Volcano yield a pre-eruptive storage depth of 4 km, which is similarto that estimated from geodetic data for the inflation source during the 2011–2012 period of caldera unrest; this supports a magmatic origin of the unrest. The level of pre-AD 726 magma storage beneathKameni was deeper than that of earlier silicic interplinian eruptions, perhaps owing tochanges in crustal stress caused by the Minoan eruption. Combined with previously published results,the melt inclusion data provide a time-integrated image of the crustal plumbing system.Mantle-derived basalts are injected into the lower crust, where they fractionate to produce evolvedmelts in bodies of hot crystal mush. Evolved residual melts separate from their parent mushes inthe 8 to >15km depth interval, then ascend rapidly into the upper crust, where they either crystallizeor accumulate as bodies of eruptible, crystal-poor magma.