You generally want to design based on C-E.
C-B means you can get to this voltage by actively pulling leakage current out of the base (failure to do so results in avalanche breakdown, which usually behaves like a zener diode, but can act very noisy or even exhibit negative resistance like an SCR, but usually very quickly like a spark discharge).
There aren't many applications where C-B would be relevant, because for example, linear applications can't draw much current and dissipate much power at those voltages anyway (due to second breakdown limitations -- see the safe operating area (SOA)), and switching applications are limited due to the slow transition between, essentially, Vcbo and Vceo (except starting from Vce(sat) and going up in turn).
Switching transistors will frequently have an RBSOA (reverse bias SOA), which shows how much voltage you can expect to drop (freely, without excessive current flow or damage) within a certain timescale after turn-off.
The reason for this phenomenon is, the transistor takes time to turn off, and as it does so, more and more of the transistor junction becomes "off". The transition from "on" to "off" occurs through the thickness of the C-B junction, so that immediately after turn-off, the region that is "off" is thin, and not much voltage can be dropped before breakdown occurs. As time goes on, the region grows, and more voltage can be withstood.
Also, avalanche introduces charge carriers, just like turning on the transistor does, so that avalanche during turn-off is just like dumping current *into* the base, when you thought you were pulling it *out* -- it acts to turn the transistor on, which not only slows turn-off time, but dissipates a heck of a lot of switching loss in the process.
So you want to be very careful that, when you turn off a BJT, you do it sharply (how fast the B-E junction turns off, controls how fast the C-B junction turns off, even after a fairly sizable time delay between those events (storage time t_stg)), and you limit the C-E voltage risetime (usually with a dV/dt snubber) to avoid this dynamic avalanche effect.
Tim