It isn't that simple. Yes, Mathematica has emphasized symbolic computation since day 1, but it has also had numerical capabilities like NIntegrate since day 1 too. There's no good reason why the symbolic capabilities of Mathematica should necessarily hurt its numerical performance. In particular, Mathematica and Matlab often use the exact same libraries for many numerical operations, e.g. Intel's MKL for many matrix operations or SuiteSparse for sparse matrices. Any inefficiencies in calling out to those libraries between Matlab and/or Mathematica are usually down to a few tweaks.

In the vast majority (90% or more) of cases the performance between Matlab and Mathematica is identical or within the margin of error. And where performance does differ substantially, Matlab isn't always coming out on top. So the question then becomes whether or not your work depends on those particular use cases. In my eyes, the Wolfram Language is fundamentally superior to Matlab. And even if Mathematica/WL lags behind Matlab is some numerical aspects, Wolfram Research will have a far easier time fix or improving that deficiency than MathWorks will have improving the poverty of symbolic computation in Matlab.

Of course, the factors in practical use cases are quite a bit more complex than what I've said above. If you're working at an engineering firm that's already knee-deep in Simulink and various Matlab Toolboxes, then the choice has been already been made for you. And sometimes it's not a question of performance but one of feature parity, e.g. Compile in the Wolfram Language vs Matlab Compiler/Runtime. Also, the complexity of the Wolfram Language is not a benefit for users looking to simply write code and get a result ASAP.