Can you get some information on how little variation in acceleration/rotation can the accelerometer/gyro measure? How often can you get a reading?
I don't think you need an 5g accelerometer because I imagine an balancing robot is falling as a rock by the time it registers 1g so 5g might be waaay over the top - this is why I believe resolution is the more intereasting parameter.
The same is probably true for the gyro: resolution and sampling rate might be more important then total number of degrees per second. Here's my reasoning: If the bot has just registered 300 degrees rotation in one second it's probably nose-down in the dirt, too late to do anything! On the other hand if the gyro allows 1kHz sampling rate, then you can get 1000 samples per second and the 300 degrees per second translate to 0.3 degrees per millisecond.
Some time in the (very) distant future I'd like to build an balancing robot myself, this is why I'm interested in anything about balancing.