It's still O(n)
(n/2
if you want to be more precise). You're splitting the work in half at each step, but it has to be recombined as you go up the chain. You cannot add all the even values without performing #even values - 1 addition operations, minimum. This just improves the odds that any given addition is with roughly equal sized operands, where looping directly would have one accumulator operand continually growing as the source operands stay the same size.
Ignoring the evenness element, think of summing all the numbers in the array this way. If you had eight elements, you'd add the even elements to the odd (four operations, leaving four elements), then add the even results to the odd results (two operations, leaving two elements), then add those two remaining values together (one operation, leaving the answer). You performed 4 + 2 + 1 == 7
operations to add 8 values. You'd have done the same number with just:
elems = [1,2,3,4,5,6,7,8]
accumulator = elems[0]
# Loop runs seven times, for all but first element, so seven additions performed
for operand in elems[1:]: # Ignore the slice cost here; avoidable, but irrelevant
accumulator += operand