In general IPv6 has a strong presumption that on-link subnets are uniformly /64, except point to point links which many folks recommend be /127 from a /64 allocation. While this is not inherently necessary, some software may be inflexible on this point and break if you use other sizes, and you will certainly confuse everyone else if you use different sizes. Personally, compared to the mess that is IPv4 subnetting, I like having all my v6 subnets the same size.
Mapped addresses of various kinds (0::/8) which won't appear on the wire do use a variety of other sizes.
ISP DHCPv6 prefix delegation to home users "ought" to be /60 or so with the home subnets being /64 on the wire, except that lots of DSL and cable modems goof up if you offer them anything larger than a single /64, so most live deployments current delegate exactly one /64.
IPv6 only cellular LTE4 data networks are tending toward delegating two /64's, one for the native v6 traffic and a differently routed one for 464xlat to tunnel single-stack v4 app traffic to a carrier nat64/dns64 gateway.
Note that there are a lot of other ways to pick 64-bit host parts of addresses besides the default EUI-64 mapping of a 48-bit MAC address:
* (unused) drop in an IEEE 64-bit MAC, if you can find one
* (unused) the mostly experimental SEND cryptographically hashed addresses (rfc3972, rfc4581, rfc4982)
* old-style pseudo-random privacy addresses, changed every 24 hours or so which are the default on Microsoft operating systems (rfc4941)
* new-style subnet-stable privacy addresses (rfc7217)
Moving out of the SLAAC realm, there are also the options of dynamic DHCPv6 pools, DHCPv6 reservations, or manually assigned host parts chosen by an administrator.
Currently you mostly run into old-style privacy, the original EUI-64 mapped, and static/manual in that order of popularity, but that is going to change.
-- Jim Leinweber
