It depends on the solver. In most cases, JuMP will reformulate (1) into (2) or (2) into (1) to best support the solver so it doesn’t matter.

But in general, modeling as a SecondOrderCone is better.

One reason is that the quadratic constraint:

@constraint(model, x^2 + y^2 <= z^2)

is actually non-convex, so it depends on JuMP or the solver figuring out that if there is that quadratic constraint, andz >= 0, then it is actually convex and equivalent to a SecondOrderCone.

(Also note that your equation (2) is not quite right. It needs <= instead of =.)