For those of us who need accurate clocks and don't have $1500 to spend, highly stable temp controlled oscillator chips are cheap and common right now. (Search eBay for OXCO)
For example, this one (which I'm using) is accurate in the PPB range:
An OXCO has excellent phase noise and short term stability, but drifts. A Cs reference has better long term stability. An Rb device has pretty good long term stability once you factor out the linear "aging" (i.e., a constant change). GPS-DO combine GPS receivers with firmware and discipline another oscillator, usually a quartz crystal because of the low phase noise.
So the right way to compare these various references is with an Allan deviation plot, which rates the stability across different time scales.
For this new Symmetricom product, they list the following Stability in their Allan deviation chart (no plot, hmmm):
Tau = 1s ADEV = 2e-10
Tau = 10s ADEV = 7e-11
Tau = 100s ADEV = 2e-11
Tau = 1000s ADEV = 7e-12
They similarly give phase noise, aging, initial calibration accuracy, etc.
FEI's 5680A Rb product at http://freqelec.com/rb_osc_fe5680a.html shows their ADEV plot, so I squinted at it and read these values:
Tau = 1s ADEV ~ 3e-11
Tau = 10s ADEV ~ 3e-12
Tau = 100s ADEV ~ 7e-13
Tau = 1000S ADEV ~ 3e-13
Further than that (days) and you start running into aging corrections.
An Rb unit you buy off eBay isn't likely to perform like this; you'd need a temperature-controlled room, and need to keep it away from 60Hz fields (the magnetic field is used to "tune" the transition, I think due to the anomalous Zeeman effect, but I could be wrong). Also, a NIST scientist told me that gravity effects are strong as well, so it must be calibrated at the same altitude as use to get these 10^-12 and better numbers.
Still, for about $100 for either an Rb or a GPS-DO, they're pretty good. The OXCO is pretty good too, but beats the Rb or GPS-DO only in phase noise. (Oh, and cost and power consumption.)