Presented at the 2014 Aquarius/SAC-D Science Team MeetingAquarius is a NASA instrument comprising three L-band (1.4 GHz) radiometers dedicated to the remote sensing of global sea surface salinity. During the nominal operation, Aquarius' radiometers are calibrated using an internal hot target. An additional empirical adjustment is performed by comparing measured and simulated antenna temperatures (Ta) over oceans. The simulation utilizes a radiative transfer model that integrates contributions from all directions around the antenna and weights them by the antenna gain patterns. The ocean calibration was used at the beginning of the mission to correct biases of a few Kelvin, and a drift of -1K during the first year of the mission. This resulted in measured Ta being within +/-0.2K of the simulations over oceans.In order to assess and improve the calibration, Aquarius' beams are pointed toward the celestial sky about once a month. These cold sky calibration (CSC) maneuvers started seven months into the mission and have been performed continuously since then. The brightness temperature (Tb) of the Sky is relatively well know and very stable in time. The Sky also has very large areas of homogenous Tb, offering flat scenes to Aquarius' relatively large footprints. The CSC is performed when the spacecraft is above the oceans in order to limit the uncertainties due to land emissivity and radio frequency interferences (more prevalent over land).During the first year of the mission, the CSC identified a bias of the order of -2K at the low end of the Tb range. A significant source of bias is the uncertainty on the antenna gain pattern. Several versions of Aquarius gain pattern model exist (derived from numerical simulations and measurements on a scale model of the spacecraft) and there is significant disagreement regarding the spillover ratio (SR). The SR is the off-Earth fraction of antenna power when the antenna is in its normal mode pointing toward the Earth. SRs differ by ~1.5% between various models, enough to create a ~1.5K/4K Ta bias over ocean/land. A first step in improving the calibration was to validate the SR as accurately as possible. For that purpose, we use measurements performed when the spacecraft was in the cold sky calibration mode (i.e. inverted) and crossing an ocean/land transition: The very large change in scene temperature makes the uncertainty on the emissivity model negligible when estimating the SR. With the new estimate of SR, the existing gain pattern is adjusted to create a hybrid model for use in the calibration.The use of the hybrid pattern results in significant improvements in the agreement between measured and simulated TA at the cold end. We will report on the CSC performed between March 2012 and September 2014 and discuss the result regarding the absolute calibration and the stability of the radiometers.