The (T) contribution makes up a relatively small part of the valence correlation energy, while its evaluation, in large basis sets and for systems with very many electrons, will dominate the CPU time. For instance, in a very recent study on SiF4[8], a CCSD(T) calculation with an AVQZ basis set on F and a VQZ+2d1f basis set on Si took 50h7' using MOLPRO on an SGI Octane workstation (768 MB of memory being allocated to the job), of which 41h30' were spent in the (T) step alone.
In addition, it was previously noted by Helgaker and coworkers[69] that the (T) contribution appears to converge faster with the basis set than the CCSD correlation energy, for which reason they actually propose its separate evaluation in a smaller basis set. In the present case, we have considered extrapolating it from AVTZ+2d1f and AVQZ+2d1f results rather than AVQZ+2d1f and AV5Z+2d1f, respectively. In our adopted notation, this becomes W[TQ;Q5;TQ5].
As seen in Table VI, the difference in quality between W[TQ;Q5;TQ5] and W[Q5;Q5;TQ5] appears to be essentially negligible. This is an important conclusion, since it means that the largest basis set calculation to be carried out is only at the CCSD level, at a fraction of the cost of the full CCSD(T) counterpart -- moreover it can be done using direct algorithms.[70]