I see your point, but I guess I don't see the difference between it being structured as a string with a second deserialization step after JSON deserialization.

I see your point, but I guess I don't see the difference between it being structured as a string with a second deserialization step after JSON deserialization. I see the convenience of a standard way to deserialize the time stamp. But either way you need multiple pieces of data for the duration to be correct and useful. Without the created-at time in your example the duration will be invalid in the presence of leap years/seconds. If you want to unambiguously encode a duration of time, it needs to be in the smallest unambiguous unit that is meaningful (usually seconds/nano seconds). That will allow your duration to be correctly used without any additional logic/metadata packed along with it.