I've seen related issues discussed here but not exactly this one. I'm following closely the current recommendations for an exome pipeline, and the GATK version,downloaded from git, was v2.5-2-gf57256b, Compiled 2013/06/06 17:28:57.
For example, I have two samples with heterozygous variants 12:81503433C>G. The AD values for the the samples in the raw vcf file, and the SNVs-only file were 15,14 and 20,15 for the two samples and these agree with what I see in IGV. There was nothing in the indels-only file at that position. The AD values were the same in the recalibrated SNVs-only file. But after combining the recalibrated SNVs and indels with CombineVariants the AD values inexplicably became 21,0 and 0,24 respectively. This seems to be happening to many variants.