Thank you Danner, I tryed but I can't get a good looking material, actually playing with refraction and transparency gave me weird results: to less it's kind of opaque, and to much is like luminous.
Other suggestions? I can of course setup a 'classic' fresnel glass, but delta and dielectric are way faster in rendering (I don't know, maybe because the calculation is done by algorithm in the node). The problem is that dielectric has some problems with Kray (I'm in trial with the demo in these days) and I'm looking for a good and faster solution.
This is another example of why I always run back to the trusty Sigma2 node. It is a very powerful node. It's how I do wax, Legos, Gummy Bears, Milk, etc. It's very flexible. Only catch is when using Radiosity you have to bypass LW11.x Unified Sampling with regard to GI and crank the GI samples to rid yourself of unwanted pre-processing artifacts. I generally start around 32 and need to go as high as 64.