The CCSD(T) method is often called the "gold standard" of computational chemistry,
because it is one of the most accurate methods applicable to reasonably large molecules.
It is particularly useful for the description of noncovalent interactions where the
inclusion of triple excitations is necessary for achieving a satisfactory accuracy.
While it is widely used as a benchmark, the accuracy of CCSD(T) interaction energies
has not been reliably quantified yet against more accurate calculations. In this work,
we compare the CCSD[T], CCSD(T), and CCSD(TQ) noniterative methods with full CCSDTQ
and CCSDT(Q) calculations. We investigate various types of noncovalent complexes [hydrogen-bonded
(water dimer, ammonia dimer, water ··· ammonia), dispersion-bound (methane dimer,
methane ··· ammonia), and π-π stacked (ethene dimer)] using various coupled-clusters
schemes up to CCSDTQ in 6-31G*(0.25), 6-31G**(0.25, 0.15), and aug-cc-pVDZ basis sets.
We show that CCSDT(Q) reproduces the CCSDTQ results almost exactly and can thus serve
as a benchmark in the cases where CCSDTQ calculations are not feasible. Surprisingly,
the CCSD[T] method provides better agreement with the benchmark values than the other
noniterative analogs, CCSD(T) and CCSD(TQ), and even than the much more expensive
iterative CCSDT scheme. The CCSD[T] interaction energies differ from the benchmark
data by less than 5 cal/mol on average (for all complexes and all basis sets), whereas
the error of CCSD(T) is 9 cal/mol. In larger systems, the difference between these
two methods can grow by as much as 0.15 kcal/mol. While this effect can be explained
only as an error compensation, the CCSD[T] method certainly deserves more attention
in accurate calculations of noncovalent interactions.