>It seems too good to be true. So what’s the catch?
You’re the decrease in solution time against the additional effort to
form and use the substructure. All that is laid out in the docs.

In particular you need an extra run to generate the substructure and any
number of stress passes to extract results. You have to balance off the
time and effort of MDOF selection, the additional generation run and the
stress passes against the time for the non-substructured run. If you’re
only interested in displacements at MDOF you don’t need stress passes,
but that’s not normally the case. In the case of spectral analysis, the
stress passes can be a huge pain in the butt when it comes time for
combining modal response. Also substructures don’t plot in POST1, so you
can’t see the displaced structure with stress passes. Same for POST26
although you can generate time-domain displacement responses for MDOF.

A potentially bigger problem is the inaccuracies inherent in matrix
condensation to form the substructure. The condensed stiffness matrix is
accurate but not the mass matrix. If you don’t carry enough MDOF or if
the MDOF are chosen improperly, the calculated response is inaccurate.

There’s a great deal of spiritual value in generating a proper
substructure and in managing and validating all the stress passes.
Substructures are mathematical abstractions which effectively lose their
geometry between the generation run and the stress pass. It takes
discipline, attention to first principles and the ability to use FEA
without the eye candy. They’re nothing like free money.