I'd bet they have very similar performance-metrics, but the yield syntax is more extensible (i.e. you're not limited to one expression) and debug-able (you can

I'd bet they have very similar performance-metrics, but the yield syntax is more extensible (i.e. you're not limited to one expression) and debug-able (you can put breakpoints within the function). Also the name and the generator is nicer (for some definition of nice): >>> def square_vals(x : list): ... return (v * v for v in x) ... >>> square_vals([1,2,3]) <generator object square_vals.<locals>.<genexpr> at 0x786b8511f5e0> >>> def square_vals_yields(x: list): ... for v in x: ... yield v * v ... >>> square_vals_yields([1,2,3]) <generator object square_vals_yields at 0x786b851f5ff0> I think it's more idiomatic to pass generator-comprehensions into functions rather than return them from functions >>> sum((v*v for v in x))