I would echo rb1957 on this one and suggest you ensure that your model has fully converged. This can be done by using an iterative process of gradual mesh refinement until your peak stresses converge on some value.
I would be careful of over refining a mesh though, as stress singularities could...