In this 2005 Notices article, Jerold Grossman tracks the proportion of papers in Math Reviews with 1, 2, 3, and >3 authors over time. His data set ends in 1999. I seem to recall reading that in 200k, for some value of k, the number of collaborative papers in math exceeded the number of singleauthored papers for the first time. But I can't find any record of this. Does anyone know what k is, what a reliable source for this kind of data is, and what the proportion of collaborative papers is now?

According to the article, the original data was provided by the AMS. I don't think that the AMS leaves this sort of data lying around on laptops on trains, so do to it again you'd have to go and ask them. I suspect that, quite reasonably, the AMS likes to know what uses their data is put to. On the other hand, data can be mined very easily from the arXiv via the API. I don't know if arXiv data would suffice for you. If so, a little scripting showed the following data for the month of October:
The ordering is by numberofauthors. So for math.KT there were 4 papers, of which 3 were single authored and 1 with 2 authors. Missing entries are 0s (so in math.NT there was a 6author paper but none with 4 or 5). So collaborations outweigh singleauthor papers by a little bit (technical term). 


I found this article from the Information Processing and Management, which may be related: "Trends in transforming scholarly communication and their implications" (2002) Unfortunately, I can't access the actual text, and am unwilling to pay for it. 


I imagine you'll find more collaborative papers in applied math than in pure math. In fact, I'd speculate that the number of authors increases steadily as you move along the continuum from most pure (e.g. category theory) to most applied (e.g. applied statistics). 

