The language specifies that time_t is an arithmetic type capable of representing times. It doesn’t require it to represent times in any particular way.
If time_t represents time as the number of seconds since some moment, the - operator will correctly compute the difference in seconds between two time_t values.
If it doesn’t (say, if the granularity is one millisecond, or if the bits of a time_t are divided into groups representing years, months, days, etc.), then the - operator can yield meaningless results.
The difftime() function, on the other hand, “knows” how a time_t represents a time, and uses that information to compute the difference in seconds.
On most implementations, simple subtraction and difftime() happen to do the same thing — but only difftime() is guaranteed to work correctly on all implementations.
Another difference: difftime() returns a result of the floating-point type double, while "-" on time_t values yields a result of type time_t. In most cases the result will be implicitly converted to the type of whatever you assign it to, but if time_t happens to be an unsigned integer type, subtraction of a later time from an earlier time will yield a very large value rather than a negative value. Every system I’ve seen implements time_t as a 32-bit or 64-bit signed integer type, but using an unsigned type is permitted — one more reason that simple subtraction of time_t values isn’t necessary meaningful.