We do know about the Px2 platform it uses 2 Pascal GPUs with 256 CUDA cores each, the only Pascal part that Nvidia currently sells with 256 CUDA cores is the GP106. You will find the GP106 was used in the GTX 1060 which currently sell for $200 each so we're looking at ~$400 for the GPUs alone (who knows what changes they make for automotive use).
$200 for the SOC is not possible given the hardware involved.
At this point, the term PX2 platform is likely a very loose thing. Tesla says they wrote their own software and was choosing between NVIDIA, AMD, and Intel. They picked the NVIDIA Titan. NVIDIA puts out a press release saying that Tesla uses the Drive PX2 platform. I suspect the real answer is that Tesla is doing things in a PX2 kind of way, but that doesn't mean they are using anything in particular that you find on a particular PX2 board as supplied by NVIDIA. SpaceX has a habit of doing this also. So likely a GP106, as it is a cut down GP104. We don't know the bin, as we don't know how fast Tesla is clocking this chip. Given that the GTX 1060 reference graphics card is expected to retail for $249, the chip cost is far, far less. I suspect around $100 to $125. Tegra's are like $40-50 in automotive. The Mobileye EyeQ3 was under $50 IIRC.