In my project (Pax et Bellum) I use thousands of grass models with alpha channel.
And I have no problem with the frame rate, because I use mipmaps and LOD.
Your grass has an alphachannel and it does not write into the depthbuffer, which means that every pixel is rendered for every grass plane every frame. And those are blended together, which means that the previous colour has to be read and mixed with the new one for every grass plane.
That's right.
But if you use mipmaps, only the smaller mipmap must be rendered when an object is far away. Rendering/blending is much faster then.
In my game I use very detailed textures (8192 x 8192 pixels). To render them as fast as possible, I use DDS textures with 9(!) Mipmaps.
Here you can see a screenshot of my game with thousands of grass models (all with alpha channel). The frame rate is higher than 60 FPS even though the level is not finished and optimized now (the grass models in this test level are old models with only 4 Mipmaps, and I use no LOD).
(click here to see the original size of the screenshot) 
) 
