Researchers tested AI on hundreds of high-value professional tasks and found models are improving—but not yet ready to do the jobs themselves.