To have more than marginal (> 10%) speedup they need a few thousand, and to have > 20% speedup they need sizes \gtrsim 10^4. This seems pretty consistent with other experiments I’ve seen over the years, e.g. with ATLAS BLAS in 2006.
See also this discussion: Matrix multiply breakthrough, AlphaTensor (could also do for other algorithms): "AlphaTensor discovered algorithms that are more efficient than the state of the art for many matrix sizes." - #7 by stevengj