In brief - anywhere you used Huffman in the past, you can use ANS instead.
ANS (or ABS) are not very useful for high end compressors. They are awkward for adaptive modeling. Even if all you care about is decode speed (so you don't mind the buffering up the models to make the encode work backwards) it's just not a big win over arithmetic coding. Things like PAQ/CM , LZMA, H264, all the high compression cases that use adaptive context models, there's no real win from ANS/ABS.
Some specific cases where I see ANS being a good win :
JPEG-ANS obviously. Won't be competitive with sophisticated coders like "packjpg" but will at least fix the cliff at low bit rate caused by Huffman coding.
JPEGNEXT-ANS. I've been thinking for a long time about writing a "JPEGNEXT". Back end coefficient classification; coefficients in each group sent by value with ANS. Front end 32x32 macroblocks with various DCT sizes. Try to keep it as simple as possible but be up to modern quality levels.
LZ-ANS. An "LZX class" (which means large window, "repeat match", entropy resets) LZ with ANS back end should be solid. Won't be quite LZMA compression levels, but way way faster.
Lossless image DPCM. ANS on prediction residual values is a clear killer. Should have a few statistics groups with block classification. No need to transmit the ANS counts if you use a parametric model ala TMW etc. Should be easy, fast, and good.
blocksort-ANS. Should replace bzip. Fast to decode.
MPEG-class video coders. Fast wavelet coders. Anywhere you are using minimal context modelling (only a few contexts) and are sending values by their magnitude (not bit plane coders).