This paper applied AVO to oil and brine discrimination for cavern carbonate reservoirs in Tarim Basin, Northwest China. The intercept (P) and gradient (G) attributes are extracted from the top and base of the cavern reservoirs encountered by 16 known wells. In the P-G cross-plot domain, oil and brine saturated cavern reservoirs are well separated, with oil wells having a positive G and brine wells having a negative G. We performed AVO forward modeling for the cavern carbonate reservoir and the results show that reservoir Vp/Vs ratio is the controlling factor in the AVO signature at the top and base of the cavern carbonate reservoir. The brine saturated case with a high Vp/Vs ratio exhibits class I AVO response, i.e., amplitude decreases with incident angle, while the oil saturated case with a low Vp/Vs ratio, exhibits class III AVO signature, i.e., amplitude increases with incident angle. Based on these findings, we designed an oil-brine classifier using the P-G attributes and implemented oil-brine prediction for 21 wells before drilling. The real drilling results show that the correct rate of this oil-brine classifier is about 81%. The results are promising and show that AVO could be an effective tool for oil and brine discrimination for cavern carbonate reservoirs.