Do LP dual variables exactly match shadow prices?

Dual variables are often interpreted as shadow prices (the change in the objective function value when a constraint is relaxed by one unit). However, it has long been pointed out that this interpretation does not always hold for degenerate linear programming problems. (e.g. https://journals.sagepub.com/doi/abs/10.3233/INF-2000-11406)

Has this issue not been resolved even in the latest version of Gurobi?

When I tested this using Gurobi 12, I found that for a small problem with three variables, the results matched almost exactly (to about 10e-14). However, for a large-scale problem with tens of millions of variables, there was a discrepancy of several percent.