Mythbusters tested this (measuring fuel consumption and efficiency in real cars), and then went back and revisited the issue after comments and concerns and re-tested. Wikipedia's final summary of their final summary:
The fundamental flaw in the MythBusters' test was that the point where the drag becomes powerful enough to inhibit a car's performance with windows down was inside their 45–55mph margin at 50mph. Going less than 50mph it is more efficient to leave the windows down, but going greater than 50mph it is more efficient to use the A/C.
ETA: It's probably a reasonable assumption that the 50mph figure will vary depending on the aerodynamic configuration of the car and the windows, as well as other differences affecting the A/C compressor's efficiency. But it's a reasonable ballpark, and reasonable cars' break-even point probably doesn't vary a whole lot.