For whatever reason I've always looked at averages rather than expected arrays. Here your average of the six stats is 12.83.
Rolling 5d6-drop-2 once for each stat gives an average of, I think, about 13.1; which in effect is +1 to two of your six listed stats.
Bit late to the party on this one, but I thought it might still be worthwhile to weigh in. Averages can be misleading, however, since it is quite rare for
every single roll to be near the average. That's why you often want to simulate a full spread of results, if possible, rather than strictly dealing with the numerical averages.
We
can use AnyDice to perform this calculation easily. Literally just took the ordinary 4d6-drop-lowest program
presented here, modified to keep the top 3 out of 5 rather than the top 3 out of 4. As you can see from the graph version (rather than the table version, which isn't very easy to make comparisons with), the expected results are {16.44, 15.21, 14.14, 13.06, 11.80, 9.93}. If we round these to the closest integer, we get {16, 15, 14, 13, 12, 10}. This is a little bit more than +1 to every stat, as the default 5e array is {15, 14, 13, 12, 10, 8}: a componentwise subtraction gives {1, 1, 1, 1, 2, 2} for a total of 8 additional points, though the biggest differences are at the lowest end of the curve (and partially caused by rounding.) The 5e point-buy exactly reproduces the default array.
Allowing a seventh roll also has a side effect of reducing the chance greatly of a character starting with a very low score, as it becomes a "dump roll". Whether this is a feature, a bug, or neither is in the eye of the beholder.
This is fair. (Incidentally,
adding a 7th roll to the above example gives {16.60, 15.48, 14.53, 13.62, 12.64, 11.45, 9.68}. Meaning that, on average, the low roll you discard is quite likely to be no less than 7, implying that all your
other rolls are usually 8+. The statistically average array becomes {17, 15, 15, 14, 13, 11, 10} with the 10 being discarded. Adding a 7th roll to the ordinary 4d6-drop-lowest option instead gives an expected average array of {16, 14, 13, 12, 11, 10, 8}, with that 14 being literally a hair's breadth from rounding up to 15 instead.)