I think that's a binomial problem … beyond that, it really depends on where your random numbers are coming from. Modern OSes have csrngs in them — and, I suppose a few shitty bits of entropy from your mouse movements, if that works at all. Does yours? Does your CPU have one of those weird intel dual-frequency hrngs? Mine has a dedicated entropy generator, but crypto nerds are quick to point out you don't need them if you have a good csrng, just a couple good seed bits and you're fine.

I'd expect what you have to repeat eventually. Use a hash, say Digest::SHA1, and I'd expect that to be pretty damn random. Feed it the time, a rand() number, and some sequential number. I'd expect that to have a lot more bits of doesn't-repeat than your method, if only because 25 bytes isn't very many when I have databases with a few hundred thousand payment records in them. Ya know?