Difference between two time_point instances is not a duration?

It does produce a duration, but there are different kinds of durations. std::chrono::duration is templatized on a representation type and a unit ratio. std::chrono::seconds for example has a unit ratio of 1, while std::chono::nanoseconds has a unit ratio of std::nano, or 1/1000000000. time points have the same template parameters.

The specific unit ratio of std::chrono::system_clock::time_point is implementation defined, but it is almost certainly less than than that of std::chrono::seconds. As such, the duration produced from subtracting those two time points has much more precision than can be represented by std::chrono::seconds. The default behaviour is to not allow assignments that lose precision with durations that have integer representations. So you can either use a duration with enough precision (std::chrono::system_clock::duration) or cast the result to the duration you want (std::chrono::duration_cast<std::chrono::seconds>(...)).

The difference between two time points is indeed a duration; but you can't implicitly convert one duration type to another, since that could silently lose precision.

If you want to reduce the precision from system_clock::duration to seconds, then you need to make the conversion explicit using a duration_cast:

delay = duration_cast<std::chrono::seconds>(t1 - t2);

Alternatively, you might want to retain the precision of the system clock:

auto delay = t1 - t2; // Probably microseconds, or nanoseconds, or something

time_point - time_point does return a duration, just not the one in the code. You could replace std::chrono::seconds with std::chrono::system_clock::duration, or you could use a duration_cast to convert to the kind you need.