It doesn't diff the versions, charges you for total storage:

http://aws.amazon.com/s3/faqs/#How_am_I_charged_for_using_Ve...

Any idea why they don't do it the same way as snapshots?

I'd guess snapshots aren't retrieved fairly often so the additional CPU and IO overhead of reconstituting a version are totally worth it given the space savings, unlike for data that's used regularly?

(Disclaimer: I have no experience in snapshot system implementation so I don't know what I'm saying)

If this works with CloudFront, it would save us a ton of effort renaming files to their MD5 hashes. We could then serve multiple concurrent versions of your website from a single file tree.

No word from them yet. They normally strip query arguments. Hopefully they'll make an exception for versioning.

There was some discussion previously, here: http://news.ycombinator.com/item?id=1067154

if only they had rsync support...

No kidding. We only push changed files to S3 after every build, but it's still a lot of bytes that go across the wire. I wish there was a way to build a new file as a diff from an old one.