Carbon dioxide and water, being present in the Earth’s mantle at concentration levels of tens to hundreds of ppm, greatly lower the peridotite solidus temperature and drastically modify the composition of produced melts. The presence of CO2 produces silica-poor, carbonate-rich liquids at the onset of melting, and these liquids shift toward silica rich compositions as the degree of melting increases. Numerous geochemical observations and experimental studies have revealed the complexity of the transition between carbonate-rich and silicate-rich melts. It is characterized by a strongly non-linear evolution and, under specific conditions, by immiscibility. To better constrain this transition, we have used the thermodynamic activity of silica as a probe of the mixing properties between molten carbonate and molten silicate. The activity of silica (image) was calculated for a large number of experimental liquids from two equilibria: olivine-orthopyroxene-melt and immiscible silicate-rich melt-carbonate-rich melt (491 data points ranging from 1 to 14 GPa and 1090 to 1800°C). We modeled image during incipient melting of the peridotite in presence of CO2 with a generalized Margules function. Our model well reproduces the silica activity–composition relationships of the experimental database, and can be used to predict the silica content of the melts coexisting with olivine and orthopyroxene. We show that water content and Ca/Mg ratio in the melts have an important influence on the image. In contrast to a recent empirical model (Dasgupta et al., 2013), the analysis of the experimental database reveals that the transition from carbonate to silicate melt with decreasing depth should occur abruptly in oceanic mantle. Our model predict that carbonatitic melts with ~ 5 wt.% SiO2 can be stabilized from ~ 150 km depth, at the onset of incipient melting by “redox melting”, up to ~ 75 km, above which the liquid evolves abruptly to a carbonated silicate composition (> ~ 25 wt.% SiO2). In the cratonic mantle lithosphere, our model predicts that carbonatitic melts are prevailing up to shallow depth, and conflicts the recent model (Russell et al., 2012) of CO2-saturation triggered by orthopyroxene assimilation during kimberlite ascent.