Given that IPV6 has a size of 1500, I would assert that carriers would not provide separate paths for IPV4 and IPV6 (they are both IP with different types), forcing them to equipment for ipv4 that would be old, redundant, more costly to maintain and less reliable. It wouldn't make any sense. Besides, doing so might easily be considered providing preferential treatment for some traffic -- a no no under rules they probably don't care much about (unless they get caught).
So 1472 should be safe for external use (though that doesn't mean an app like DNS that doesn't know about EDNS will accept it), and if you are talking internal nets, you can more likely know your network layout in which case jumbo packet sizes apply for for non-fragmented packets so for 4096 - 4068 bytes, and for intel's cards with 9014 byte buffers, a package size of ... wait...8086 bytes, would be the max...coincidence? snicker
****UPDATE****
Various answers give maximum values allowed by 1 SW vendor or various answers assuming encapsulation. The user didn't ask for the lowest value possible (like "0" for a safe UDP size), but the largest safe packet size.
Encapsulation values for various layers can be included multiple times. Since once you've encapsulated a stream -- there is nothing prohibiting, say, a VPN layer below that and a complete duplication of encapsulation layers above that.
Since the question was about maximum safe values, I'm assuming that they are talking about the maximum safe value for a UDP packet that can be received. Since no UDP packet is guaranteed, if you receive a UDP packet, the largest safe size would be 1 packet over IPv4 or 1472 bytes.
Note -- if you are using IPv6, the maximum size would be 1452 bytes, as IPv6's header size is 40 bytes vs. IPv4's 20 byte size (and either way, one must still allow 8 bytes for the UDP header).