how to interpret interpret timeit command in Python

You didn't include the -n argument to %timeit in your second example, so ipython varies the number of repetitions based on how long a trial-run takes; the faster the piece of code being tested, the more iterations are done to get a more accurate per-iteration time value.

Moreover, the tests are run several times to try and minimise external factors (for example, when your OS just happened to schedule a disk buffer flush and everything else becomes a little bit slower). This is where the 'best of 3' comes in; the tests were run 3 times in a row and the best timings were picked.

See the %timeit magic command documentation, which includes these options and their default behaviour:

-n<N>: execute the given statement <N> times in a loop. If this value is not given, a fitting value is chosen.

-r<R>: repeat the loop iteration <R> times and take the best result. Default: 3

Your first example does use -n 10 so it was run just 10 times.

Because creating a generator object with a generator expression is near-instant, ipython can execute the loop way more often than executing a list comprehension (which has to execute the for loop and produce a list object with all the results there and then). Remember that a generator expression does not do any work until you drive iteration.

If you wanted to compare how long a generator expression takes to produce the same results as a list comprehension, you'd have to actually iterate. You could pass the expression to a list() call to actually produce a list too:

%timeit -n 10 myGen = (item for item in L if item < 15); list(myGen)

This'll be slower as a generator has a little more overhead than a list comprehension:

In [1]: L=[-13, -24, -21, -3, -23, -15, -14, -27, -13, -12]

In [2]: %timeit -n 10 myList = [item for item in L if item < 15]
10 loops, best of 3: 1.29 µs per loop

In [3]: %timeit -n 10 myGen = (item for item in L if item < 15); list(myGen)
10 loops, best of 3: 1.72 µs per loop

Note that you have to re-create the generator each test iteration because generators can produce their output just once.