There is really no universally accepted cut-off, and I believe that even reasonable cutoffs change by application. In other words, it depends on what you are going to do with the data, and it depends on the subject matter as well. EXAMPLE: In spectroscopy, many different X variables can have correlation of 0.95 and everyone is okay with this.
Another way to look at this is the variance inflation factor (in PROC REG) which indicates how the collinearity will affect the variance of the predicted values. Most people set 3 as a cutoff for this factor, but I don't think this is based on any hard science.
--
Paige Miller